1Software College of Northeastern University Shenyang Liaoning 110819 China2Department of Radiology Chinese PLA General Hospital Shenyang 110015 China
Correspondence should be addressed to Huiyan Jiang jianghyswcneueducn
Received 4 September 2016 Revised 18 November 2016 Accepted 8 December 2016
Academic Editor Enzo Terreno
Copyright copy 2016 Qingjiao Sun et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
Pathological image enhancement is a significant topic in the field of pathological image processing This paper proposes a highdynamic range (HDR) pathological image enhancement method based on improved bias field correction and guided image filter(GIF) Firstly a preprocessing including stain normalization and wavelet denoising is performed for Haematoxylin and Eosin (Hand E) stained pathological imageThen an improved bias field correctionmodel is developed to enhance the influence of light forhigh-frequency part in image and correct the intensity inhomogeneity and detail discontinuity of image Next HDR pathologicalimage is generated based on least square method using low dynamic range (LDR) image H and E channel images Finally thefine enhanced image is acquired after the detail enhancement process Experiments with 140 pathological images demonstrate theperformance advantages of our proposed method as compared with related work
1 Introduction
Pathological image is an important basis for computer aideddiagnosis and is regarded as the gold standard in diseasediagnosis Cell segmentation and identification are criticalsteps in various medical diagnoses it is difficult to acquireaccurate cell segmentation results because of low contrastand noise of image To address this issue it is necessary toenhance pathological image before cell segmentation processThe enhancement of pathological image can improve imagequality and contrast and it can provide more objective andreliable data support for doctors This is of great significanceand strong application value as it improves detection effi-ciency andmedical diagnostic accuracymeanwhile reducingmanual diagnosis error and human costs
There are a variety of image enhancement methods andframeworks to improve image quality Traditional histogramequalization- (HE-) based methods [1ndash3] are widely usedfor image enhancement owing to their simplicity Besidesmany new models and algorithms are proposed to processimage Reference [4] proposed an efficient transformation-
free approach for color image enhancement which manip-ulates pixels value directly in source RGB color spaceReference [5] presented a color image enhancement methodusing daubechies wavelet transform and HIS color spaceReference [6] enhanced the pathological anatomy imagesbased on superresolution to improve medical diagnosis Inrecent years the Retinex theory is widely used in the medicalimage processing [7] Many algorithms based on Retinextheory such as single-scale Retinex (SSR) [8] multiscaleRetinex (MSR) [9] McCann99 [10] and Frackle-McCann[11] have been developed for image enhancement Thereare also many other methods to enhance high dynamicrange (HDR) images [12ndash14] Reference [15] put forwarda Bilateral Filtering-Dynamic Range Partitioning (BF-DRP)algorithm which can use bilateral filter (BF) on image toextract a coarse component and a details component whichare processed independently and then recombined togetherto obtain final enhanced image Reference [16] improved theBF-DRP algorithm by adding an adaptive Gaussian filter tosmooth the imbalanced variation of gradient Reference [17]proposed guided image filter (GIF) and [18] applied that in
Hindawi Publishing CorporationBioMed Research InternationalVolume 2016 Article ID 7478219 11 pageshttpdxdoiorg10115520167478219
2 BioMed Research International
HDR image denoising and enhancement which can avoidgradient flipping artifacts
However all these methods mainly target natural imagesor infrared images the effect is unsatisfactory when they areused on pathological images It remains a challenging taskto obtain a good enhancement result for color pathologicalimages because of three main issues intensity inhomogeneityof image detail discontinuity in tissue structures and lowerdynamic range of image To address those issues in this paperwe propose a HDR pathological image enhancement methodbased on GIF and improved bias field correction model Themain contributions of this paper are as follows First a newpathological image operational process is designed to denoiseand enhance the source lowdynamic range (LDR) image Sec-ond an improved bias field correction model is proposed tocorrect the intensity inhomogeneity and detail discontinuityof imageThird a newmethod to generate HDR pathologicalimage is presented using LDR image and Haematoxylin (H)and Eosin (E) channel image after stain separation
The remainder of the paper is organized as follows InSection 2 we introduce the related work Our proposedpathological image enhancement method is described inSection 3 Section 4 shows our experiments results andcompares them with other enhancement methods Finallyconclusions are summarized in Section 5
2 Related Work
21 Bias FieldCorrectionModel The intensity inhomogeneityis a common phenomenon of medical images which isattributed to many factors such as nonuniform illuminationimaging equipment defect and the complexity of humantissues Intensity inhomogeneity is a critical factor that affectssome image processing because it will influence the trueintensity region of different tissues and then lead to the errorsof image segmentation or other image analysis processesTherefore it is one necessary step to remove the intensityinhomogeneity from the image The bias field is a popularmathematical assumption of image intensity inhomogeneitywhich is generally accepted at present and it manifested asthe smoothly varying of intensity within the same tissue of theimage This assumption can be represented by the followingmathematical model [19]
119868119900 = 119861 times 119869 + 119899 (1)
where 119868119900 is the observed image with intensity inhomogeneity119869 is the true image 119861 is the bias field and 119899 is Gaussian noisewith zero mean which can be ignored after denoising processand then get 119868119900 = 119861 times 119869 There are usually two assumptionsfor the above model [20]
(1) The bias field 119861 is smoothly varying That is the biasfield approximates a constant in a small neighbor-hood of every pixel in the observed image
(2) The true image 119869 describes the physical property oftissues and the value of this property should be thesame in the same tissue Thus we assume that thepixel value within every tissue of the true image 119869 isa constant
The bias field correction procedure is used to remove thebias field from image and finally to obtain the corrected image119868119900119861 There are many bias field correction methods and[20] proposed a new algorithm called multiplicative intrinsiccomponent optimization (MICO) for bias field estimationwhich achieved a good result In that paper the bias field isrepresented by a linear combination of a group of smoothbasis functions 1198921 119892119872 as follows
119861 (119909) = 119882119879119866 (119909) (2)
where 119866(119909) = (1198921(119909) 119892119872(119909))119879 is a column vectorvalued function composed of basis functions (sdot)119879 is thetranspose operator and 119909 is pixel point in the image and119882 = (1205961 120596119872)119879 is a column vector of the coefficientsTherefore bias field estimation can be viewed as to find theoptimal coefficients 1205961 120596119872 of the linear combination119861(119909) = sum119872119896=1 120596119896119892119896
In addition the true image is formulated as the followingmodel
119869 (119909) =119873
sum119894=1
119888119894120583119894 (119909) (3)
where 119888119894 is the constant of the 119894th tissue and there are119873 tissuesin the image 120583119894(119909) is the percentage of the 119894th tissue being inthe pixel 119909 there are 120583119894(119909) = 1 for 119909 isin Ω119894 and 120583119894(119909) = 0 for119909 notin Ω119894 and Ω119894 is the range of 119894th tissue
In order to calculate the bias field 119861 that paper proposedan energy minimization function as follows
119865 (119861 119869) = intΩ
10038161003816100381610038161198680 (119909) minus 119861 (119909) 119869 (119909)10038161003816100381610038162 119889119909 (4)
whereΩ is the whole image domain Aftermerging equations(2) (3) and (4) the energy function 119865 can be expressed as
We can see that the energy 119865 is the function about variables120583 119888 120596 The minimization of 119865 can be achieved by alternatelysolving one variable with the other two fixed And we canfinally obtain the bias field corrected image after solving thebias field estimation
22 Guided Image Filter GIF filters the input image byconsidering the guidance image It is a smoothing operatorwhich can smooth filtering preserve the edge details andavoid the artifacts effectively It is fast and easy to implementand can obtain a nice visual quality GIF is derived from alocal linear transformation model considering the content ofa guidance image The filtering process at a pixel 119894 can beformulated as follows [17]
Image correction based on improved biasfield model
Generate HDR pathological image
Detail enhancement for HDR pathologicalimage based on GIF
Pathological image
Enhanced pathological image
Figure 1 The flowchart of proposed pathological image enhance-ment method
where 119868 is the guidance image and the 119902 is the linear transformof 119868 in window 120596119896 centered at pixel 119896 (119886119896 119887119896) are the locallinear coefficients the calculating formula of which is asfollows
119886119896 =(1 |120596|) sum119894isin120596119896 119868119894119901119894 minus 120583119896119901119896
1205902119896+ 120576
119887119896 = 119901119896 minus 119886119896120583119896(7)
where 120583119896 and 1205902119896 are the mean value and variance separatelyof image 119868 in window 120596119896 |120596| is the number of pixels within120596119896 119901 is the input image and 119901119896 is the mean of 119901 in window120596119896 120576 is the regularization parameter also called smooth factorand used to prevent 119886119896 being too large
However a pixel 119894 is involved in more than one windowthat covers 119894 when computing (2) and 119902119894 has different valuein different windows Therefore we can get the final 119902119894 valueby averaging all 119902119894 and thus the output of the filter can berepresented as follows
Here the average local linear coefficients are 119886119894 =(1|120596|)sum119896isin120596119894 119886119896 and 119887119894 = (1|120596|)sum119896isin120596119894 1198871198963 Proposed Pathological Image
Enhancement Method
As shown in Figure 1 the generic process of proposedpathological image enhancement method is introduced
31 Image Preprocessing There are always some undesirableproblems such as hypochromasia hyperchromasia and colorvariations in the pathological images acquired from stainedtissue due to different staining solutions of manufactures
different scanners or microscopes and different stainingprotocols of labs which can block the image observationand image interpretation [21] To improve image quality weuse Reinhardrsquos stain normalization method [22] to bring thepathological image into a better color appearance of a targetimage selected by pathologists
There is much additive noise in the pathological imageafter stain normalization thus we adopt denoising methodto smooth the image Some denoising methods often cannottake into account both denoising and preserving image detailand lead to edge blur problem To solve this problem thewavelet denoising method [23] is performed which canpreserve the image edge and other feature information whiledenoising
32 Improved Bias Field Correction Model The traditionalbias fieldmodel concerns the intensity inhomogeneity causedby bias field in the low frequency part but ignores the detaildiscontinuity of high-frequency part caused by different lightTherefore we improve the bias field model through includingthe detail discontinuity
119868119900 = 119861 times 119863 times 119869 + 119899 (9)
where119863 is the factor of detail discontinuityWe have the following assumptions about factor119863(1) 119863 is mainly high-frequency information and only
exists in the edge region of image(2) 119863 is slowly varying in the edge region and has no
significant gradient variation
After denoising (9) becomes 119868119900 = 119861 times 119863 times 119869 At thebase of MICO algorithm [20] we use high-pass filter toremove low frequency information and then superimposehigh-frequency information to enhance image details and toremove detail discontinuity
33 Generate HDR Pathological Image The dynamic rangeof pathological image acquired by digital image acquisitiondevices is very limited Generally the luminance range of dig-ital image described by RGB color model is about two ordersof magnitude (256) and there are always some overexposureand underexposure parts in the image However the dynamicrange of a real scene is much wider than that of the image andthe range that can be observed by human eyes is also muchlarger In medical diagnosis the dynamic range of sensorscan reach 104 even 105 but LDR image cannot record allinformation On the contrary HDR image can record bothRGB information and actual luminance information of pixelsand candescribe image details betterWe could provide betterimages with clearer details and more reliable information forsubsequent cell segmentation and identification if using LDRimages to generate HDR image It is noted that HDR imagesneed to be compressed because they cannot be displayedon traditional screen directly which is the tone mappingoperator (TMO) of HDR image [24] Therefore HDR imageenhancement can change the dynamic range of luminance oforiginal image meanwhile acquiring more details of the realscene
4 BioMed Research International
Table 1 Comparison of GIF and BF in run time (unit second)
Methods ImagesGroup 1 (1280 times 960) Group 2 (640 times 480) Group 3 (320 times 240) Group 4 (100 times 100) Group 5 (50 times 50)
We usually need many image sequences with differentexposure of the same scene to generate HDR image But inthis paper we generate HDR image sequences based on stainseparationmethod [25] In this case we cannot determine theexposure time Here one sample strategy is that we estimatethe exposure time by calculating the ratio of mean values ofthe sequential images
The steps of generating HDR image are as follows
(1) Separate the corrected pathology image into H chan-nel stained byHaematoxylin and E channel stained byEosin based on the stain separation method
(2) Estimate the ratio of exposure time of adjacent imagesbased on their means of pixels for both correctedpathology images and H and E channel images Thenfit curve of pixel relationship based on least squaremethod
(3) Calculate camera response function based on Mit-sunaga algorithm [26] by replacing the pixel valuewith pixel relationship curve and then generate HDRpathological image
34 Detail Enhancement Based on GIF The enhancementmethod based on BF [27] can reduce the loss of image detailinformation but it has gradient reverse problem In this paperwe replace BFwithGIF becauseGIF can avoid image gradientreverse and reduce calculation cost In order to prove thesuperiority of GIF we design five groups of images withdifferent sizes And there are 10 color pathological imageswith the same size in each group Every image is processedwith BF and GIF respectively and the run time is recordedsuccessively Finally we calculate the average run time forevery group The comparison result is shown in Table 1
It is clear that GIF takes much less run time than BFwhenprocessing the same image Therefore in our image enhance-ment method we firstly transform the HDR pathology imageinto YCbCr color space (the type of sampling is 4 4 4) thendivide the luminance component into base layer and detaillayer by GIF rather than BF next compress the dynamic rangefor base layer using histogram mapping to reduce noise andenhance detail information meanwhile enhance the detaillayer by adaptive masking and finally combine the enhancedluminance component and color component together Thebasic procedure is shown in Figure 2
The equation of hierarchical processing for image basedon GIF is shown as follows
where 119865119866 is the GIF 119868119884 is the logarithm of luminance 119884 aftertransforming the HDR image into YCbCr color space 119868119861 isthe base layer of image including the grey information of tex-ture region in image and 119868119863 is the detail layer including high-frequency detail information such as edges (119909 119910) describethe coordinate of the pixel
The result of GIF is related to the radius 119903 of filter windowand the regular parameter (smooth factor) 120576 Figure 3 showshow these two parameters affect the hierarchical resultsfor pathology image Figure 3(a) is the original pathologicalimage and (b) is the Y channel image by transforming theoriginal image into YCbCr color space (c) (e) and (g) showthe base layer with 119903 = 6 120576 = 004 119903 = 10 120576 = 001and 119903 = 10 120576 = 004 respectively (d) (f) and (h) showthe detail layer of image corresponding to (c) (e) and (g)respectively It is seen that the detail layer with larger windowradius and regular parameter can obtain much more detailinformation
4 Experiments and Comparison
In this paper all the experiments are implemented usingMatlab R2014a development tool in the 64-bit Windows 7operating system (8-core CPU 340GHz 8 G memory) andour experimental data is actual clinical pathological imageswhich are RGB color images stained by Haematoxylin andEosin (H and E) The picture format is TIFF and spatialresolution is 1280 lowast 960 The data includes 123 liver tissueimages and 17 lung tissue images that is 140 images intotal Besides the parameters of the GIF are 119903 = 10 120576 =004 The enhancement result of our proposed method isshown in Figure 4 Figure 4(a) is the original pathologicalimage and Figure 4(b) is the reference image which has goodstain quality and is used in the stain normalization processFigure 4(c) is the normalized image after stain normalizationFigure 4(d) is the result after denoising and correcting byimproved bias field model Figures 4(e) and 4(f) show theE channel and H channel images after stain separation ofFigure 4(d) respectively Figure 4(g) is the HDR pathologicalimage generated with Figures 4(c) 4(e) and 4(f) displayedby TMO Figure 4(h) is the final detail enhanced HDRpathological image
It is observed that our proposed method can obtain agood image enhancement effect in Figure 4This method canimprove image luminance and contrast meanwhile it canpreserve the image detail structure well In order to verifythe effectiveness of our algorithm we compare our methodwith various Retinex-based methods and some HE-basedalgorithms namely Frankle-McCann Retinex [11] SSR [28]MSR [29] Dualistic Sub-Image Histogram Equalization
BioMed Research International 5
Transform into YCbCr space
Color component (CbCr)
Guided image filter
Base layer Detail layer
Histogram mapping Adaptive enhancement
Base layer in LDR Detail layer in LDR
Detail enhanced LDR luminance component
Merge
HDR pathological image
LDR image
Logarithm fetch on luminance component (Y)
Figure 2 The flowchart of detail enhancement based on GIF
(a) Original image (b) Y channel image (c) Base layer image (119903 = 6 120576 = 004) (d) Detail layer image (119903 = 6 120576 =004)
(e) Base layer image (119903 = 10 120576 =001)
(f) Detail layer image (119903 = 10 120576 =001)
(g) Base layer image (119903 = 10 120576 =004)
(h) Detail layer image (119903 = 10 120576 =004)
Figure 3 Illustration of layer separation on pathological image with different parameters
(DSIHE) [30] Minimum Mean Brightness Error Bi-Histo-gram Equalization (MMBEBHE) [31] Recursive Mean-Separate Histogram Equalization (RMSHE) [32] and Recur-sive Sub-Image Histogram Equalization (RSIHE) [33] Allthese methods are applied to our 140 pathological images
successively andwefinally get 140 groups of enhanced imagesThere are 9 images in every group one original image and 8result images enhanced by 8 image enhancement algorithmsFigure 5 demonstrates one of the groups where Figure 5(a) isthe original image Figure 5(b)ndashFigure 5(h) are the enhanced
6 BioMed Research International
(a) Original image (b) Reference image for stain normalization
(c) Normalized image (d) Corrected image by improved bias field model
(e) E channel image (f) H channel image
(g) HDR image after tone mapping (h) Detail enhanced HDR image
Figure 4 The result of our proposed HDR pathological image enhancement method
image by the above comparison algorithms respectively andFigure 5(i) is our result
We can see from Figure 5 that all the above methodscan enhance the pathological images to some extent But the
differences between the results are a little big To evaluatecorrectly the enhancement results of different algorithmsthis paper analyze the corresponding images from both thesubjective and objective aspects
BioMed Research International 7
(a) Original image (b) Frankle-McCann Retinex (c) SSR
(d) MSR (e) DSIHE (f) MMBEBHE
(g) RMSHE (h) RSIHE (i) Proposed
Figure 5 The comparison of our method with other different image enhancement methods
Table 2 The number of best enhanced image selected by pathologists for different algorithms
41 Subjective Evaluation Pathological image is the goldstandard in disease diagnosis In order to verify the prac-ticability of our algorithm we extract 50 groups of imagesrandomly and invite three pathologists to choose one bestimage from every group The selected image should be theoptimal in visual quality and the most helpful to clinicalapplication from the perspective of a pathologist The finalresult is shown in Table 2
As Table 2 shows the results of HE-based algorithmsare basically not recognized by pathologists because there isserious distortion problem in the enhanced image of thosealgorithms Most of the detail information in those images islost and the cell regions cannot be distinguished Although itis possible to obtain the optimal enhancement effect for everyone of the other algorithms our proposed method is betterstatistically having the maximum amount of the selected best
8 BioMed Research International
Table 3 Data comparison of our method with other different image enhancement methods
Algorithms EvaluationsPSNR (dB) SD Mean EME Entropy (bit) Run time (second)
image So from a pathologistrsquos viewpoint our pathologicalimage enhancement method can improve the image qualitya lot
42 Objective Evaluation On the other hand we also analyzethe experiment results quantitatively in five well-known met-rics namely peak signal noise ratio (PSNR) [34] standarddeviation (SD) mean measure of enhancement (EME) [35]and information entropy [34]
PSNR is widely used to evaluate the quality of imageThehigher PSNR value denotes that the image could suppressnoise better The calculation equation of PSNR is as follows
and where 119883 and 119884 are input image and output imagerespectively119873 is the total number of image pixels and119871 is thedynamic range of pixel values The unit of PSNR is decibels(dB)
Themean value is used to evaluate the average luminanceof image The higher mean value represents that the image isbrighter SD is also a popular metric in image enhancementand is used to estimate the contrast of image The imagecontrast is grater if the SD value is higher For mean and SDwe compute the average value of three channels of the colorpathological image in our experiments
EME is one well-known blind-reference image qualityassessment (IQA)metric It gives a quality score to each imagebased on the image contrast The larger EME value representsthe more detail information and more obvious variation inlocal region The definition of EME is as follows
EME11989611198962 =1
1198961 11989621198962
sum119897=1
1198961
sum119896=1
20 log 119868120596max119896119897
119868120596min119896119897 (13)
where the test image is divided into 1198961 times 1198962 small blocksand 119868120596max119896119897 and 119868120596min119896119897 represent the maximum andminimumvalues of pixel respectively in block 120596119896119897
Information entropy is an important metric to measurethe content of image And the higher value indicates an imagewith richer details The equation of entropy is as follows
119867 = minus119872
sum119896=1
119901119896log2119901119896 (14)
where119872 is the gray levels of image and 119901119896 is the probabilityof gray level 119896 in the whole image
All the objective evaluation result data is shown in Table 3and all the data is an average value of our 140 pathologicalimages We can see from Table 3 that all these algorithmscould enhance image to some extent The SD values of HE-based methods are much higher than other methods Buttheir contribution is not outstanding when taking the seriousdistortion problem of those enhanced images into accountIn addition the PSNR value of our algorithm is not as goodas the Retinex-based algorithms but we obtain the bestresults in all the other metrics of mean EME and entropyAnd considering that we just use the original image andthe enhanced image to replace the noise-free image and testimage in the definition of PSNR the contribution of PSNR islimited So in conclusion our method is superior to others
Moreover in order to indicate the complexity of ouralgorithm we compare the run times of all the algorithmsrunning in the same platform and using the same imageas described above The final average result is shown in thelast column of Table 3 It is obvious that our algorithm ismuch faster than Retinex-based algorithms Although therun time of HE-based algorithm is the least the time cost ofour algorithm is worthy in view of the quality of enhancedimages
43 Cell Segmentation We also use a simple automaticcell segmentation method based on morphology to ver-ify that our image enhancement algorithm can improvesubsequent image segmentation and other image analysisprocesses In this cell segmentation method firstly transformthe original RGB image into gray-scale image and thenconduct image opening operation image reconstructionimage binarization image erosion and dilation operationsand image denoising [36] We select 20 groups of enhanced
BioMed Research International 9
(a) Original image (b) Ground truth (c) Frankle-McCann (d) SSR (e) MSR
images in which the cell regions are relatively obvious to dotest Figure 6 demonstrates one group segmentation resultwhere Figure 6(a) is the original RGB pathological imageFigure 6(b) is the ground truth segmented manually underthe guidance of pathologists and Figure 6(c)ndashFigure 6(j)are obtained from segmenting the corresponding enhancedimages
It is clear that the cell segmentation result of ouralgorithm is more close to the ground truth In order tocompare the segmentation results quantitatively we adopttwo standard segmentation metrics to evaluate namely dicecoefficient [37] and Jaccard index [38] The definitions ofthem are as follows
DC (SRTR) = 2 timesNum (pixelSR cap pixelTR)Num (pixelSR) +Num (pixelTR)
JI (SRTR) = Num (pixelSR cap pixelTR)Num (pixelSR cup pixelTR)
(15)
where SR is the segmentation region TR is the true regionof the target and Num(pixel) is the number of relevantpixel points The average results of performance metricsare displayed in Table 4 We can see that the result ofour proposed method is the best Our HDR pathologicalimage enhancement method is superior to the comparisonalgorithms as we can improve the image segmentation andpathological analysis better
In sum we proposed that HDR pathological imageenhancement method obtains a better result according to
both pathologistsrsquo subjective evaluation and quantitativeanalysis in data and the cell segmentationmethod also provesthe better performance of our method as the quality anddetail of original image are both improved
5 Conclusions
This paper proposes new HDR pathological image enhance-ment methods based on GIF and improved bias fieldcorrection model First stain normalization and waveletdenoising operations are used in image preprocessing Andthe improved bias field model is introduced to correct theintensity inhomogeneity and detail discontinuity of imageThen the HDR pathological image is generated using LDRimage and H and E channel images Next the 119884 componentof HDR image is separated into base layer and detail layerby GIF and the two layers are enhanced separately Finallythe fine enhanced image is acquired after combining the 119884component and the color components To verify the effective-ness of the proposed method we perform the enhancementexperiments using 140 pathological images The experimentresults and comparisons with related work demonstrate thatour proposed method improves the image quality in terms ofhuman vision PSNR SD mean EME information entropyand cell segmentation
Competing Interests
The authors declare that there were no competing interestsregarding the publication of this article
10 BioMed Research International
Acknowledgments
This work was supported by the National Natural ScienceFoundation of China (no 61472073)
References
[1] SM Pizer E P Amburn J D Austin et al ldquoAdaptive histogramequalization and its variationsrdquo Computer Vision Graphics ampImage Processing vol 39 no 3 pp 355ndash368 1987
[2] V Vani and K V M Prashanth ldquoColor image enhancementtechniques in Wireless Capsule Endoscopyrdquo in Proceedings ofthe IEEE International Conference on Trends in AutomationCommunications and Computing Technology (I-TACT rsquo15) vol1 pp 1ndash6 Bangalore India December 2015
[3] H Cao L Tian J Liu H Wang and S Feng ldquoColor imageenhancement using power-constraint histogram equalizationfor AMOLEDrdquo in Proceedings of the IEEE 11th InternationalConference on ASIC (ASICON rsquo15) pp 1ndash4 IEEE ChengduChina November 2015
[4] N M Kwok G Fang and H Y Shi ldquoColor enhancementfor images from digital camera using a transformation-freeapproachrdquo in Proceedings of the 9th International Conferenceon Sensing Technology (ICST rsquo15) pp 168ndash172 IEEE AucklandNew Zealand December 2015
[5] S D Nikam and R U Yawale ldquoColor image enhancementusing daubechies wavelet transform and HIS color modelrdquoin Proceedings of the International Conference on IndustrialInstrumentation and Control (ICIC rsquo15) pp 1323ndash1327 IEEEPune India May 2015
[6] L G Villanueva G M Callico F Tobajas et al ldquoMedicaldiagnosis improvement through image quality enhancementbased on super-resolutionrdquo in Proceedings of the 13th EuromicroConference onDigital SystemDesign ArchitecturesMethods andTools (DSD rsquo10) pp 259ndash262 IEEE Lille France September2010
[7] W Sun F Li and Q Zhang ldquoThe applications of improvedretinex algorithm for X-ray medical image enhancementrdquo inProceedings of the International Conference on Computer Scienceand Service System (CSSS rsquo12) pp 1655ndash1658 IEEE NanjingChina August 2012
[8] G Zhang D Sun P Yan H Zhao and Z Li ldquoA LDCT imagecontrast enhancement algorithm based on single-scale retinextheoryrdquo in Proceedings of the International Conference on Com-putational Intelligence for Modelling Control amp Automation pp1282ndash1287 IEEE Computer Society Vienna Austria December2008
[9] S Setty N K Srinath and M C Hanumantharaju ldquoDevel-opment of multiscale retinex algorithm for medical imageenhancement based on multi-rate samplingrdquo in Proceedingsof the International Conference on Signal Processing ImageProcessing amp Pattern Recognition pp 145ndash150 2013
[10] J Mccann ldquoLessons learned from mondrians applied to realimages and color gamutsrdquo in Proceedings of the Color andImaging Conference vol 8 pp 1ndash8 1999
[11] B V Funt F Ciurea and J J McCann ldquoRetinex in Matlabrdquo inProceedings of the Color and Imaging Conference pp 112ndash121Scottsdale Ariz USA November 2000
[12] K Kim J Bae and J Kim ldquoNatural hdr image tone mappingbased on retinexrdquo IEEE Transactions on Consumer Electronicsvol 57 no 4 pp 1807ndash1814 2011
[13] J Kuang G M Johnson and M D Fairchild ldquoiCAM06 arefined image appearance model for HDR image renderingrdquoJournal of Visual Communication amp Image Representation vol18 no 5 pp 406ndash414 2007
[14] M-L Song H-Q Wang C Chen X-Q Ye and W-K GuldquoTone mapping for high dynamic range image using a proba-bilistic modelrdquo Journal of Software vol 20 no 3 pp 734ndash7432010
[15] F Branchitta M Diani G Corsini and M Romagnoli ldquoNewtechnique for the visualization of high dynamic range infraredimagesrdquo Optical Engineering vol 48 no 9 Article ID 0964012009
[16] C Zuo ldquoDisplay and detail enhancement for high-dynamic-range infrared imagesrdquo Optical Engineering vol 50 no 12Article ID 127401 pp 895ndash900 2011
[17] K He J Sun and X Tang ldquoGuided image filteringrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol35 no 6 pp 1397ndash1409 2013
[18] N Liu and D Zhao ldquoDetail enhancement for high-dynamic-range infrared images based on guided image filterrdquo InfraredPhysics amp Technology vol 67 pp 138ndash147 2014
[19] C Li R Huang Z Ding et al ldquoA level set method for imagesegmentation in the presence of intensity inhomogeneities withapplication toMRIrdquo IEEE Transactions on Image Processing vol20 no 7 pp 2007ndash2016 2011
[20] C Li J C Gore and C Davatzikos ldquoMultiplicative intrinsiccomponent optimization (MICO) for MRI bias field estimationand tissue segmentationrdquoMagnetic Resonance Imaging vol 32no 7 pp 913ndash923 2014
[21] A Vahadane T Peng A Sethi et al ldquoStructure-preservingcolor normalization and sparse stain separation for histologicalimagesrdquo IEEE Transactions on Medical Imaging vol 35 no 8pp 1962ndash1971 2016
[22] E Reinhard M Ashikhmin B Gooch and P Shirley ldquoColortransfer between imagesrdquo IEEE Computer Graphics amp Applica-tions vol 21 no 5 pp 34ndash41 2001
[23] D L Donoho ldquoDe-noising by soft-thresholdingrdquo IEEE Trans-actions on Information Theory vol 41 no 3 pp 613ndash627 1995
[24] E Zhang H Yang and M Xu ldquoA novel tone mappingmethod for high dynamic range image by incorporating edge-preserving filter into method based on retinexrdquo Applied Mathe-matics amp Information Sciences vol 9 no 1 pp 411ndash417 2015
[25] M Macenko M Niethammer J S Marron et al ldquoA methodfor normalizing histology slides for quantitative analysisrdquo inProceedings of the IEEE International Conference on Symposiumon Biomedical Imaging From Nano To Macro pp 1107ndash1110IEEE Press 2009
[26] T Mitsunaga and S K Nayar ldquoRadiometric self calibrationrdquoin Proceedings of the IEEE Computer Society Conference onComputer Vision and Pattern Recognition vol 1 p 1374 FortCollins Colo USA June 1999
[27] F Durand and J Dorsey ldquoFast bilateral filtering for the displayof high-dynamic-range imagesrdquoACMTransactions onGraphicsvol 21 no 3 pp 257ndash266 2002
[28] D J Jobson Z-U Rahman andG AWoodell ldquoProperties andperformance of a centersurround retinexrdquo IEEE Transactionson Image Processing vol 6 no 3 pp 451ndash462 1997
[29] D J Jobson Z-U Rahman and G A Woodell ldquoA multiscaleretinex for bridging the gap between color images and thehuman observation of scenesrdquo IEEE Transactions on ImageProcessing vol 6 no 7 pp 965ndash976 1997
BioMed Research International 11
[30] YWangQ Chen andB Zhang ldquoImage enhancement based onequal area dualistic sub-image histogram equalizationmethodrdquoIEEE Transactions on Consumer Electronics vol 45 no 1 pp68ndash75 1999
[31] S-D Chen and A R Ramli ldquoMinimum mean brightnesserror bi-histogram equalization in contrast enhancementrdquo IEEETransactions on Consumer Electronics vol 49 no 4 pp 1310ndash1319 2003
[32] S-D Chen and A R Ramli ldquoContrast enhancement usingrecursive mean-separate histogram equalization for scalablebrightness preservationrdquo IEEE Transactions on Consumer Elec-tronics vol 49 no 4 pp 1301ndash1309 2003
[33] K S Sim C P Tso and Y Y Tan ldquoRecursive sub-imagehistogram equalization applied to gray scale imagesrdquo PatternRecognition Letters vol 28 no 10 pp 1209ndash1221 2007
[34] V L Jaya and R Gopikakumari ldquoIEM a new image enhance-ment metric for contrast and sharpness measurementsrdquo Inter-national Journal of Computer Applications vol 79 no 9 pp 1ndash92013
[35] S S Agaian K Panetta and A M Grigoryan ldquoTransform-based image enhancement algorithms with performance mea-surerdquo IEEE Transactions on Image Processing vol 10 no 3 pp367ndash382 2001
[36] M Braiki A Benzinou K Nasreddine S Labidi and NHymery ldquoSegmentation of dendritic cells from microscopicimages using mathematical morphologyrdquo in Proceedings ofthe 2nd International Conference on Advanced Technologies forSignal and Image Processing (ATSIP rsquo16) pp 282ndash287 MonastirTunisia March 2016
[37] L R Dice ldquoMeasures of the amount of ecologic associationbetween speciesrdquo Ecology vol 26 no 3 pp 297ndash302 1945
[38] P Jaccard ldquoThe distribution of the flora in the alpine zonerdquoNewPhytologist vol 11 no 2 pp 37ndash50 1912
HDR image denoising and enhancement which can avoidgradient flipping artifacts
However all these methods mainly target natural imagesor infrared images the effect is unsatisfactory when they areused on pathological images It remains a challenging taskto obtain a good enhancement result for color pathologicalimages because of three main issues intensity inhomogeneityof image detail discontinuity in tissue structures and lowerdynamic range of image To address those issues in this paperwe propose a HDR pathological image enhancement methodbased on GIF and improved bias field correction model Themain contributions of this paper are as follows First a newpathological image operational process is designed to denoiseand enhance the source lowdynamic range (LDR) image Sec-ond an improved bias field correction model is proposed tocorrect the intensity inhomogeneity and detail discontinuityof imageThird a newmethod to generate HDR pathologicalimage is presented using LDR image and Haematoxylin (H)and Eosin (E) channel image after stain separation
The remainder of the paper is organized as follows InSection 2 we introduce the related work Our proposedpathological image enhancement method is described inSection 3 Section 4 shows our experiments results andcompares them with other enhancement methods Finallyconclusions are summarized in Section 5
2 Related Work
21 Bias FieldCorrectionModel The intensity inhomogeneityis a common phenomenon of medical images which isattributed to many factors such as nonuniform illuminationimaging equipment defect and the complexity of humantissues Intensity inhomogeneity is a critical factor that affectssome image processing because it will influence the trueintensity region of different tissues and then lead to the errorsof image segmentation or other image analysis processesTherefore it is one necessary step to remove the intensityinhomogeneity from the image The bias field is a popularmathematical assumption of image intensity inhomogeneitywhich is generally accepted at present and it manifested asthe smoothly varying of intensity within the same tissue of theimage This assumption can be represented by the followingmathematical model [19]
119868119900 = 119861 times 119869 + 119899 (1)
where 119868119900 is the observed image with intensity inhomogeneity119869 is the true image 119861 is the bias field and 119899 is Gaussian noisewith zero mean which can be ignored after denoising processand then get 119868119900 = 119861 times 119869 There are usually two assumptionsfor the above model [20]
(1) The bias field 119861 is smoothly varying That is the biasfield approximates a constant in a small neighbor-hood of every pixel in the observed image
(2) The true image 119869 describes the physical property oftissues and the value of this property should be thesame in the same tissue Thus we assume that thepixel value within every tissue of the true image 119869 isa constant
The bias field correction procedure is used to remove thebias field from image and finally to obtain the corrected image119868119900119861 There are many bias field correction methods and[20] proposed a new algorithm called multiplicative intrinsiccomponent optimization (MICO) for bias field estimationwhich achieved a good result In that paper the bias field isrepresented by a linear combination of a group of smoothbasis functions 1198921 119892119872 as follows
119861 (119909) = 119882119879119866 (119909) (2)
where 119866(119909) = (1198921(119909) 119892119872(119909))119879 is a column vectorvalued function composed of basis functions (sdot)119879 is thetranspose operator and 119909 is pixel point in the image and119882 = (1205961 120596119872)119879 is a column vector of the coefficientsTherefore bias field estimation can be viewed as to find theoptimal coefficients 1205961 120596119872 of the linear combination119861(119909) = sum119872119896=1 120596119896119892119896
In addition the true image is formulated as the followingmodel
119869 (119909) =119873
sum119894=1
119888119894120583119894 (119909) (3)
where 119888119894 is the constant of the 119894th tissue and there are119873 tissuesin the image 120583119894(119909) is the percentage of the 119894th tissue being inthe pixel 119909 there are 120583119894(119909) = 1 for 119909 isin Ω119894 and 120583119894(119909) = 0 for119909 notin Ω119894 and Ω119894 is the range of 119894th tissue
In order to calculate the bias field 119861 that paper proposedan energy minimization function as follows
119865 (119861 119869) = intΩ
10038161003816100381610038161198680 (119909) minus 119861 (119909) 119869 (119909)10038161003816100381610038162 119889119909 (4)
whereΩ is the whole image domain Aftermerging equations(2) (3) and (4) the energy function 119865 can be expressed as
We can see that the energy 119865 is the function about variables120583 119888 120596 The minimization of 119865 can be achieved by alternatelysolving one variable with the other two fixed And we canfinally obtain the bias field corrected image after solving thebias field estimation
22 Guided Image Filter GIF filters the input image byconsidering the guidance image It is a smoothing operatorwhich can smooth filtering preserve the edge details andavoid the artifacts effectively It is fast and easy to implementand can obtain a nice visual quality GIF is derived from alocal linear transformation model considering the content ofa guidance image The filtering process at a pixel 119894 can beformulated as follows [17]
Image correction based on improved biasfield model
Generate HDR pathological image
Detail enhancement for HDR pathologicalimage based on GIF
Pathological image
Enhanced pathological image
Figure 1 The flowchart of proposed pathological image enhance-ment method
where 119868 is the guidance image and the 119902 is the linear transformof 119868 in window 120596119896 centered at pixel 119896 (119886119896 119887119896) are the locallinear coefficients the calculating formula of which is asfollows
119886119896 =(1 |120596|) sum119894isin120596119896 119868119894119901119894 minus 120583119896119901119896
1205902119896+ 120576
119887119896 = 119901119896 minus 119886119896120583119896(7)
where 120583119896 and 1205902119896 are the mean value and variance separatelyof image 119868 in window 120596119896 |120596| is the number of pixels within120596119896 119901 is the input image and 119901119896 is the mean of 119901 in window120596119896 120576 is the regularization parameter also called smooth factorand used to prevent 119886119896 being too large
However a pixel 119894 is involved in more than one windowthat covers 119894 when computing (2) and 119902119894 has different valuein different windows Therefore we can get the final 119902119894 valueby averaging all 119902119894 and thus the output of the filter can berepresented as follows
Here the average local linear coefficients are 119886119894 =(1|120596|)sum119896isin120596119894 119886119896 and 119887119894 = (1|120596|)sum119896isin120596119894 1198871198963 Proposed Pathological Image
Enhancement Method
As shown in Figure 1 the generic process of proposedpathological image enhancement method is introduced
31 Image Preprocessing There are always some undesirableproblems such as hypochromasia hyperchromasia and colorvariations in the pathological images acquired from stainedtissue due to different staining solutions of manufactures
different scanners or microscopes and different stainingprotocols of labs which can block the image observationand image interpretation [21] To improve image quality weuse Reinhardrsquos stain normalization method [22] to bring thepathological image into a better color appearance of a targetimage selected by pathologists
There is much additive noise in the pathological imageafter stain normalization thus we adopt denoising methodto smooth the image Some denoising methods often cannottake into account both denoising and preserving image detailand lead to edge blur problem To solve this problem thewavelet denoising method [23] is performed which canpreserve the image edge and other feature information whiledenoising
32 Improved Bias Field Correction Model The traditionalbias fieldmodel concerns the intensity inhomogeneity causedby bias field in the low frequency part but ignores the detaildiscontinuity of high-frequency part caused by different lightTherefore we improve the bias field model through includingthe detail discontinuity
119868119900 = 119861 times 119863 times 119869 + 119899 (9)
where119863 is the factor of detail discontinuityWe have the following assumptions about factor119863(1) 119863 is mainly high-frequency information and only
exists in the edge region of image(2) 119863 is slowly varying in the edge region and has no
significant gradient variation
After denoising (9) becomes 119868119900 = 119861 times 119863 times 119869 At thebase of MICO algorithm [20] we use high-pass filter toremove low frequency information and then superimposehigh-frequency information to enhance image details and toremove detail discontinuity
33 Generate HDR Pathological Image The dynamic rangeof pathological image acquired by digital image acquisitiondevices is very limited Generally the luminance range of dig-ital image described by RGB color model is about two ordersof magnitude (256) and there are always some overexposureand underexposure parts in the image However the dynamicrange of a real scene is much wider than that of the image andthe range that can be observed by human eyes is also muchlarger In medical diagnosis the dynamic range of sensorscan reach 104 even 105 but LDR image cannot record allinformation On the contrary HDR image can record bothRGB information and actual luminance information of pixelsand candescribe image details betterWe could provide betterimages with clearer details and more reliable information forsubsequent cell segmentation and identification if using LDRimages to generate HDR image It is noted that HDR imagesneed to be compressed because they cannot be displayedon traditional screen directly which is the tone mappingoperator (TMO) of HDR image [24] Therefore HDR imageenhancement can change the dynamic range of luminance oforiginal image meanwhile acquiring more details of the realscene
4 BioMed Research International
Table 1 Comparison of GIF and BF in run time (unit second)
Methods ImagesGroup 1 (1280 times 960) Group 2 (640 times 480) Group 3 (320 times 240) Group 4 (100 times 100) Group 5 (50 times 50)
We usually need many image sequences with differentexposure of the same scene to generate HDR image But inthis paper we generate HDR image sequences based on stainseparationmethod [25] In this case we cannot determine theexposure time Here one sample strategy is that we estimatethe exposure time by calculating the ratio of mean values ofthe sequential images
The steps of generating HDR image are as follows
(1) Separate the corrected pathology image into H chan-nel stained byHaematoxylin and E channel stained byEosin based on the stain separation method
(2) Estimate the ratio of exposure time of adjacent imagesbased on their means of pixels for both correctedpathology images and H and E channel images Thenfit curve of pixel relationship based on least squaremethod
(3) Calculate camera response function based on Mit-sunaga algorithm [26] by replacing the pixel valuewith pixel relationship curve and then generate HDRpathological image
34 Detail Enhancement Based on GIF The enhancementmethod based on BF [27] can reduce the loss of image detailinformation but it has gradient reverse problem In this paperwe replace BFwithGIF becauseGIF can avoid image gradientreverse and reduce calculation cost In order to prove thesuperiority of GIF we design five groups of images withdifferent sizes And there are 10 color pathological imageswith the same size in each group Every image is processedwith BF and GIF respectively and the run time is recordedsuccessively Finally we calculate the average run time forevery group The comparison result is shown in Table 1
It is clear that GIF takes much less run time than BFwhenprocessing the same image Therefore in our image enhance-ment method we firstly transform the HDR pathology imageinto YCbCr color space (the type of sampling is 4 4 4) thendivide the luminance component into base layer and detaillayer by GIF rather than BF next compress the dynamic rangefor base layer using histogram mapping to reduce noise andenhance detail information meanwhile enhance the detaillayer by adaptive masking and finally combine the enhancedluminance component and color component together Thebasic procedure is shown in Figure 2
The equation of hierarchical processing for image basedon GIF is shown as follows
where 119865119866 is the GIF 119868119884 is the logarithm of luminance 119884 aftertransforming the HDR image into YCbCr color space 119868119861 isthe base layer of image including the grey information of tex-ture region in image and 119868119863 is the detail layer including high-frequency detail information such as edges (119909 119910) describethe coordinate of the pixel
The result of GIF is related to the radius 119903 of filter windowand the regular parameter (smooth factor) 120576 Figure 3 showshow these two parameters affect the hierarchical resultsfor pathology image Figure 3(a) is the original pathologicalimage and (b) is the Y channel image by transforming theoriginal image into YCbCr color space (c) (e) and (g) showthe base layer with 119903 = 6 120576 = 004 119903 = 10 120576 = 001and 119903 = 10 120576 = 004 respectively (d) (f) and (h) showthe detail layer of image corresponding to (c) (e) and (g)respectively It is seen that the detail layer with larger windowradius and regular parameter can obtain much more detailinformation
4 Experiments and Comparison
In this paper all the experiments are implemented usingMatlab R2014a development tool in the 64-bit Windows 7operating system (8-core CPU 340GHz 8 G memory) andour experimental data is actual clinical pathological imageswhich are RGB color images stained by Haematoxylin andEosin (H and E) The picture format is TIFF and spatialresolution is 1280 lowast 960 The data includes 123 liver tissueimages and 17 lung tissue images that is 140 images intotal Besides the parameters of the GIF are 119903 = 10 120576 =004 The enhancement result of our proposed method isshown in Figure 4 Figure 4(a) is the original pathologicalimage and Figure 4(b) is the reference image which has goodstain quality and is used in the stain normalization processFigure 4(c) is the normalized image after stain normalizationFigure 4(d) is the result after denoising and correcting byimproved bias field model Figures 4(e) and 4(f) show theE channel and H channel images after stain separation ofFigure 4(d) respectively Figure 4(g) is the HDR pathologicalimage generated with Figures 4(c) 4(e) and 4(f) displayedby TMO Figure 4(h) is the final detail enhanced HDRpathological image
It is observed that our proposed method can obtain agood image enhancement effect in Figure 4This method canimprove image luminance and contrast meanwhile it canpreserve the image detail structure well In order to verifythe effectiveness of our algorithm we compare our methodwith various Retinex-based methods and some HE-basedalgorithms namely Frankle-McCann Retinex [11] SSR [28]MSR [29] Dualistic Sub-Image Histogram Equalization
BioMed Research International 5
Transform into YCbCr space
Color component (CbCr)
Guided image filter
Base layer Detail layer
Histogram mapping Adaptive enhancement
Base layer in LDR Detail layer in LDR
Detail enhanced LDR luminance component
Merge
HDR pathological image
LDR image
Logarithm fetch on luminance component (Y)
Figure 2 The flowchart of detail enhancement based on GIF
(a) Original image (b) Y channel image (c) Base layer image (119903 = 6 120576 = 004) (d) Detail layer image (119903 = 6 120576 =004)
(e) Base layer image (119903 = 10 120576 =001)
(f) Detail layer image (119903 = 10 120576 =001)
(g) Base layer image (119903 = 10 120576 =004)
(h) Detail layer image (119903 = 10 120576 =004)
Figure 3 Illustration of layer separation on pathological image with different parameters
(DSIHE) [30] Minimum Mean Brightness Error Bi-Histo-gram Equalization (MMBEBHE) [31] Recursive Mean-Separate Histogram Equalization (RMSHE) [32] and Recur-sive Sub-Image Histogram Equalization (RSIHE) [33] Allthese methods are applied to our 140 pathological images
successively andwefinally get 140 groups of enhanced imagesThere are 9 images in every group one original image and 8result images enhanced by 8 image enhancement algorithmsFigure 5 demonstrates one of the groups where Figure 5(a) isthe original image Figure 5(b)ndashFigure 5(h) are the enhanced
6 BioMed Research International
(a) Original image (b) Reference image for stain normalization
(c) Normalized image (d) Corrected image by improved bias field model
(e) E channel image (f) H channel image
(g) HDR image after tone mapping (h) Detail enhanced HDR image
Figure 4 The result of our proposed HDR pathological image enhancement method
image by the above comparison algorithms respectively andFigure 5(i) is our result
We can see from Figure 5 that all the above methodscan enhance the pathological images to some extent But the
differences between the results are a little big To evaluatecorrectly the enhancement results of different algorithmsthis paper analyze the corresponding images from both thesubjective and objective aspects
BioMed Research International 7
(a) Original image (b) Frankle-McCann Retinex (c) SSR
(d) MSR (e) DSIHE (f) MMBEBHE
(g) RMSHE (h) RSIHE (i) Proposed
Figure 5 The comparison of our method with other different image enhancement methods
Table 2 The number of best enhanced image selected by pathologists for different algorithms
41 Subjective Evaluation Pathological image is the goldstandard in disease diagnosis In order to verify the prac-ticability of our algorithm we extract 50 groups of imagesrandomly and invite three pathologists to choose one bestimage from every group The selected image should be theoptimal in visual quality and the most helpful to clinicalapplication from the perspective of a pathologist The finalresult is shown in Table 2
As Table 2 shows the results of HE-based algorithmsare basically not recognized by pathologists because there isserious distortion problem in the enhanced image of thosealgorithms Most of the detail information in those images islost and the cell regions cannot be distinguished Although itis possible to obtain the optimal enhancement effect for everyone of the other algorithms our proposed method is betterstatistically having the maximum amount of the selected best
8 BioMed Research International
Table 3 Data comparison of our method with other different image enhancement methods
Algorithms EvaluationsPSNR (dB) SD Mean EME Entropy (bit) Run time (second)
image So from a pathologistrsquos viewpoint our pathologicalimage enhancement method can improve the image qualitya lot
42 Objective Evaluation On the other hand we also analyzethe experiment results quantitatively in five well-known met-rics namely peak signal noise ratio (PSNR) [34] standarddeviation (SD) mean measure of enhancement (EME) [35]and information entropy [34]
PSNR is widely used to evaluate the quality of imageThehigher PSNR value denotes that the image could suppressnoise better The calculation equation of PSNR is as follows
and where 119883 and 119884 are input image and output imagerespectively119873 is the total number of image pixels and119871 is thedynamic range of pixel values The unit of PSNR is decibels(dB)
Themean value is used to evaluate the average luminanceof image The higher mean value represents that the image isbrighter SD is also a popular metric in image enhancementand is used to estimate the contrast of image The imagecontrast is grater if the SD value is higher For mean and SDwe compute the average value of three channels of the colorpathological image in our experiments
EME is one well-known blind-reference image qualityassessment (IQA)metric It gives a quality score to each imagebased on the image contrast The larger EME value representsthe more detail information and more obvious variation inlocal region The definition of EME is as follows
EME11989611198962 =1
1198961 11989621198962
sum119897=1
1198961
sum119896=1
20 log 119868120596max119896119897
119868120596min119896119897 (13)
where the test image is divided into 1198961 times 1198962 small blocksand 119868120596max119896119897 and 119868120596min119896119897 represent the maximum andminimumvalues of pixel respectively in block 120596119896119897
Information entropy is an important metric to measurethe content of image And the higher value indicates an imagewith richer details The equation of entropy is as follows
119867 = minus119872
sum119896=1
119901119896log2119901119896 (14)
where119872 is the gray levels of image and 119901119896 is the probabilityof gray level 119896 in the whole image
All the objective evaluation result data is shown in Table 3and all the data is an average value of our 140 pathologicalimages We can see from Table 3 that all these algorithmscould enhance image to some extent The SD values of HE-based methods are much higher than other methods Buttheir contribution is not outstanding when taking the seriousdistortion problem of those enhanced images into accountIn addition the PSNR value of our algorithm is not as goodas the Retinex-based algorithms but we obtain the bestresults in all the other metrics of mean EME and entropyAnd considering that we just use the original image andthe enhanced image to replace the noise-free image and testimage in the definition of PSNR the contribution of PSNR islimited So in conclusion our method is superior to others
Moreover in order to indicate the complexity of ouralgorithm we compare the run times of all the algorithmsrunning in the same platform and using the same imageas described above The final average result is shown in thelast column of Table 3 It is obvious that our algorithm ismuch faster than Retinex-based algorithms Although therun time of HE-based algorithm is the least the time cost ofour algorithm is worthy in view of the quality of enhancedimages
43 Cell Segmentation We also use a simple automaticcell segmentation method based on morphology to ver-ify that our image enhancement algorithm can improvesubsequent image segmentation and other image analysisprocesses In this cell segmentation method firstly transformthe original RGB image into gray-scale image and thenconduct image opening operation image reconstructionimage binarization image erosion and dilation operationsand image denoising [36] We select 20 groups of enhanced
BioMed Research International 9
(a) Original image (b) Ground truth (c) Frankle-McCann (d) SSR (e) MSR
images in which the cell regions are relatively obvious to dotest Figure 6 demonstrates one group segmentation resultwhere Figure 6(a) is the original RGB pathological imageFigure 6(b) is the ground truth segmented manually underthe guidance of pathologists and Figure 6(c)ndashFigure 6(j)are obtained from segmenting the corresponding enhancedimages
It is clear that the cell segmentation result of ouralgorithm is more close to the ground truth In order tocompare the segmentation results quantitatively we adopttwo standard segmentation metrics to evaluate namely dicecoefficient [37] and Jaccard index [38] The definitions ofthem are as follows
DC (SRTR) = 2 timesNum (pixelSR cap pixelTR)Num (pixelSR) +Num (pixelTR)
JI (SRTR) = Num (pixelSR cap pixelTR)Num (pixelSR cup pixelTR)
(15)
where SR is the segmentation region TR is the true regionof the target and Num(pixel) is the number of relevantpixel points The average results of performance metricsare displayed in Table 4 We can see that the result ofour proposed method is the best Our HDR pathologicalimage enhancement method is superior to the comparisonalgorithms as we can improve the image segmentation andpathological analysis better
In sum we proposed that HDR pathological imageenhancement method obtains a better result according to
both pathologistsrsquo subjective evaluation and quantitativeanalysis in data and the cell segmentationmethod also provesthe better performance of our method as the quality anddetail of original image are both improved
5 Conclusions
This paper proposes new HDR pathological image enhance-ment methods based on GIF and improved bias fieldcorrection model First stain normalization and waveletdenoising operations are used in image preprocessing Andthe improved bias field model is introduced to correct theintensity inhomogeneity and detail discontinuity of imageThen the HDR pathological image is generated using LDRimage and H and E channel images Next the 119884 componentof HDR image is separated into base layer and detail layerby GIF and the two layers are enhanced separately Finallythe fine enhanced image is acquired after combining the 119884component and the color components To verify the effective-ness of the proposed method we perform the enhancementexperiments using 140 pathological images The experimentresults and comparisons with related work demonstrate thatour proposed method improves the image quality in terms ofhuman vision PSNR SD mean EME information entropyand cell segmentation
Competing Interests
The authors declare that there were no competing interestsregarding the publication of this article
10 BioMed Research International
Acknowledgments
This work was supported by the National Natural ScienceFoundation of China (no 61472073)
References
[1] SM Pizer E P Amburn J D Austin et al ldquoAdaptive histogramequalization and its variationsrdquo Computer Vision Graphics ampImage Processing vol 39 no 3 pp 355ndash368 1987
[2] V Vani and K V M Prashanth ldquoColor image enhancementtechniques in Wireless Capsule Endoscopyrdquo in Proceedings ofthe IEEE International Conference on Trends in AutomationCommunications and Computing Technology (I-TACT rsquo15) vol1 pp 1ndash6 Bangalore India December 2015
[3] H Cao L Tian J Liu H Wang and S Feng ldquoColor imageenhancement using power-constraint histogram equalizationfor AMOLEDrdquo in Proceedings of the IEEE 11th InternationalConference on ASIC (ASICON rsquo15) pp 1ndash4 IEEE ChengduChina November 2015
[4] N M Kwok G Fang and H Y Shi ldquoColor enhancementfor images from digital camera using a transformation-freeapproachrdquo in Proceedings of the 9th International Conferenceon Sensing Technology (ICST rsquo15) pp 168ndash172 IEEE AucklandNew Zealand December 2015
[5] S D Nikam and R U Yawale ldquoColor image enhancementusing daubechies wavelet transform and HIS color modelrdquoin Proceedings of the International Conference on IndustrialInstrumentation and Control (ICIC rsquo15) pp 1323ndash1327 IEEEPune India May 2015
[6] L G Villanueva G M Callico F Tobajas et al ldquoMedicaldiagnosis improvement through image quality enhancementbased on super-resolutionrdquo in Proceedings of the 13th EuromicroConference onDigital SystemDesign ArchitecturesMethods andTools (DSD rsquo10) pp 259ndash262 IEEE Lille France September2010
[7] W Sun F Li and Q Zhang ldquoThe applications of improvedretinex algorithm for X-ray medical image enhancementrdquo inProceedings of the International Conference on Computer Scienceand Service System (CSSS rsquo12) pp 1655ndash1658 IEEE NanjingChina August 2012
[8] G Zhang D Sun P Yan H Zhao and Z Li ldquoA LDCT imagecontrast enhancement algorithm based on single-scale retinextheoryrdquo in Proceedings of the International Conference on Com-putational Intelligence for Modelling Control amp Automation pp1282ndash1287 IEEE Computer Society Vienna Austria December2008
[9] S Setty N K Srinath and M C Hanumantharaju ldquoDevel-opment of multiscale retinex algorithm for medical imageenhancement based on multi-rate samplingrdquo in Proceedingsof the International Conference on Signal Processing ImageProcessing amp Pattern Recognition pp 145ndash150 2013
[10] J Mccann ldquoLessons learned from mondrians applied to realimages and color gamutsrdquo in Proceedings of the Color andImaging Conference vol 8 pp 1ndash8 1999
[11] B V Funt F Ciurea and J J McCann ldquoRetinex in Matlabrdquo inProceedings of the Color and Imaging Conference pp 112ndash121Scottsdale Ariz USA November 2000
[12] K Kim J Bae and J Kim ldquoNatural hdr image tone mappingbased on retinexrdquo IEEE Transactions on Consumer Electronicsvol 57 no 4 pp 1807ndash1814 2011
[13] J Kuang G M Johnson and M D Fairchild ldquoiCAM06 arefined image appearance model for HDR image renderingrdquoJournal of Visual Communication amp Image Representation vol18 no 5 pp 406ndash414 2007
[14] M-L Song H-Q Wang C Chen X-Q Ye and W-K GuldquoTone mapping for high dynamic range image using a proba-bilistic modelrdquo Journal of Software vol 20 no 3 pp 734ndash7432010
[15] F Branchitta M Diani G Corsini and M Romagnoli ldquoNewtechnique for the visualization of high dynamic range infraredimagesrdquo Optical Engineering vol 48 no 9 Article ID 0964012009
[16] C Zuo ldquoDisplay and detail enhancement for high-dynamic-range infrared imagesrdquo Optical Engineering vol 50 no 12Article ID 127401 pp 895ndash900 2011
[17] K He J Sun and X Tang ldquoGuided image filteringrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol35 no 6 pp 1397ndash1409 2013
[18] N Liu and D Zhao ldquoDetail enhancement for high-dynamic-range infrared images based on guided image filterrdquo InfraredPhysics amp Technology vol 67 pp 138ndash147 2014
[19] C Li R Huang Z Ding et al ldquoA level set method for imagesegmentation in the presence of intensity inhomogeneities withapplication toMRIrdquo IEEE Transactions on Image Processing vol20 no 7 pp 2007ndash2016 2011
[20] C Li J C Gore and C Davatzikos ldquoMultiplicative intrinsiccomponent optimization (MICO) for MRI bias field estimationand tissue segmentationrdquoMagnetic Resonance Imaging vol 32no 7 pp 913ndash923 2014
[21] A Vahadane T Peng A Sethi et al ldquoStructure-preservingcolor normalization and sparse stain separation for histologicalimagesrdquo IEEE Transactions on Medical Imaging vol 35 no 8pp 1962ndash1971 2016
[22] E Reinhard M Ashikhmin B Gooch and P Shirley ldquoColortransfer between imagesrdquo IEEE Computer Graphics amp Applica-tions vol 21 no 5 pp 34ndash41 2001
[23] D L Donoho ldquoDe-noising by soft-thresholdingrdquo IEEE Trans-actions on Information Theory vol 41 no 3 pp 613ndash627 1995
[24] E Zhang H Yang and M Xu ldquoA novel tone mappingmethod for high dynamic range image by incorporating edge-preserving filter into method based on retinexrdquo Applied Mathe-matics amp Information Sciences vol 9 no 1 pp 411ndash417 2015
[25] M Macenko M Niethammer J S Marron et al ldquoA methodfor normalizing histology slides for quantitative analysisrdquo inProceedings of the IEEE International Conference on Symposiumon Biomedical Imaging From Nano To Macro pp 1107ndash1110IEEE Press 2009
[26] T Mitsunaga and S K Nayar ldquoRadiometric self calibrationrdquoin Proceedings of the IEEE Computer Society Conference onComputer Vision and Pattern Recognition vol 1 p 1374 FortCollins Colo USA June 1999
[27] F Durand and J Dorsey ldquoFast bilateral filtering for the displayof high-dynamic-range imagesrdquoACMTransactions onGraphicsvol 21 no 3 pp 257ndash266 2002
[28] D J Jobson Z-U Rahman andG AWoodell ldquoProperties andperformance of a centersurround retinexrdquo IEEE Transactionson Image Processing vol 6 no 3 pp 451ndash462 1997
[29] D J Jobson Z-U Rahman and G A Woodell ldquoA multiscaleretinex for bridging the gap between color images and thehuman observation of scenesrdquo IEEE Transactions on ImageProcessing vol 6 no 7 pp 965ndash976 1997
BioMed Research International 11
[30] YWangQ Chen andB Zhang ldquoImage enhancement based onequal area dualistic sub-image histogram equalizationmethodrdquoIEEE Transactions on Consumer Electronics vol 45 no 1 pp68ndash75 1999
[31] S-D Chen and A R Ramli ldquoMinimum mean brightnesserror bi-histogram equalization in contrast enhancementrdquo IEEETransactions on Consumer Electronics vol 49 no 4 pp 1310ndash1319 2003
[32] S-D Chen and A R Ramli ldquoContrast enhancement usingrecursive mean-separate histogram equalization for scalablebrightness preservationrdquo IEEE Transactions on Consumer Elec-tronics vol 49 no 4 pp 1301ndash1309 2003
[33] K S Sim C P Tso and Y Y Tan ldquoRecursive sub-imagehistogram equalization applied to gray scale imagesrdquo PatternRecognition Letters vol 28 no 10 pp 1209ndash1221 2007
[34] V L Jaya and R Gopikakumari ldquoIEM a new image enhance-ment metric for contrast and sharpness measurementsrdquo Inter-national Journal of Computer Applications vol 79 no 9 pp 1ndash92013
[35] S S Agaian K Panetta and A M Grigoryan ldquoTransform-based image enhancement algorithms with performance mea-surerdquo IEEE Transactions on Image Processing vol 10 no 3 pp367ndash382 2001
[36] M Braiki A Benzinou K Nasreddine S Labidi and NHymery ldquoSegmentation of dendritic cells from microscopicimages using mathematical morphologyrdquo in Proceedings ofthe 2nd International Conference on Advanced Technologies forSignal and Image Processing (ATSIP rsquo16) pp 282ndash287 MonastirTunisia March 2016
[37] L R Dice ldquoMeasures of the amount of ecologic associationbetween speciesrdquo Ecology vol 26 no 3 pp 297ndash302 1945
[38] P Jaccard ldquoThe distribution of the flora in the alpine zonerdquoNewPhytologist vol 11 no 2 pp 37ndash50 1912
Image correction based on improved biasfield model
Generate HDR pathological image
Detail enhancement for HDR pathologicalimage based on GIF
Pathological image
Enhanced pathological image
Figure 1 The flowchart of proposed pathological image enhance-ment method
where 119868 is the guidance image and the 119902 is the linear transformof 119868 in window 120596119896 centered at pixel 119896 (119886119896 119887119896) are the locallinear coefficients the calculating formula of which is asfollows
119886119896 =(1 |120596|) sum119894isin120596119896 119868119894119901119894 minus 120583119896119901119896
1205902119896+ 120576
119887119896 = 119901119896 minus 119886119896120583119896(7)
where 120583119896 and 1205902119896 are the mean value and variance separatelyof image 119868 in window 120596119896 |120596| is the number of pixels within120596119896 119901 is the input image and 119901119896 is the mean of 119901 in window120596119896 120576 is the regularization parameter also called smooth factorand used to prevent 119886119896 being too large
However a pixel 119894 is involved in more than one windowthat covers 119894 when computing (2) and 119902119894 has different valuein different windows Therefore we can get the final 119902119894 valueby averaging all 119902119894 and thus the output of the filter can berepresented as follows
Here the average local linear coefficients are 119886119894 =(1|120596|)sum119896isin120596119894 119886119896 and 119887119894 = (1|120596|)sum119896isin120596119894 1198871198963 Proposed Pathological Image
Enhancement Method
As shown in Figure 1 the generic process of proposedpathological image enhancement method is introduced
31 Image Preprocessing There are always some undesirableproblems such as hypochromasia hyperchromasia and colorvariations in the pathological images acquired from stainedtissue due to different staining solutions of manufactures
different scanners or microscopes and different stainingprotocols of labs which can block the image observationand image interpretation [21] To improve image quality weuse Reinhardrsquos stain normalization method [22] to bring thepathological image into a better color appearance of a targetimage selected by pathologists
There is much additive noise in the pathological imageafter stain normalization thus we adopt denoising methodto smooth the image Some denoising methods often cannottake into account both denoising and preserving image detailand lead to edge blur problem To solve this problem thewavelet denoising method [23] is performed which canpreserve the image edge and other feature information whiledenoising
32 Improved Bias Field Correction Model The traditionalbias fieldmodel concerns the intensity inhomogeneity causedby bias field in the low frequency part but ignores the detaildiscontinuity of high-frequency part caused by different lightTherefore we improve the bias field model through includingthe detail discontinuity
119868119900 = 119861 times 119863 times 119869 + 119899 (9)
where119863 is the factor of detail discontinuityWe have the following assumptions about factor119863(1) 119863 is mainly high-frequency information and only
exists in the edge region of image(2) 119863 is slowly varying in the edge region and has no
significant gradient variation
After denoising (9) becomes 119868119900 = 119861 times 119863 times 119869 At thebase of MICO algorithm [20] we use high-pass filter toremove low frequency information and then superimposehigh-frequency information to enhance image details and toremove detail discontinuity
33 Generate HDR Pathological Image The dynamic rangeof pathological image acquired by digital image acquisitiondevices is very limited Generally the luminance range of dig-ital image described by RGB color model is about two ordersof magnitude (256) and there are always some overexposureand underexposure parts in the image However the dynamicrange of a real scene is much wider than that of the image andthe range that can be observed by human eyes is also muchlarger In medical diagnosis the dynamic range of sensorscan reach 104 even 105 but LDR image cannot record allinformation On the contrary HDR image can record bothRGB information and actual luminance information of pixelsand candescribe image details betterWe could provide betterimages with clearer details and more reliable information forsubsequent cell segmentation and identification if using LDRimages to generate HDR image It is noted that HDR imagesneed to be compressed because they cannot be displayedon traditional screen directly which is the tone mappingoperator (TMO) of HDR image [24] Therefore HDR imageenhancement can change the dynamic range of luminance oforiginal image meanwhile acquiring more details of the realscene
4 BioMed Research International
Table 1 Comparison of GIF and BF in run time (unit second)
Methods ImagesGroup 1 (1280 times 960) Group 2 (640 times 480) Group 3 (320 times 240) Group 4 (100 times 100) Group 5 (50 times 50)
We usually need many image sequences with differentexposure of the same scene to generate HDR image But inthis paper we generate HDR image sequences based on stainseparationmethod [25] In this case we cannot determine theexposure time Here one sample strategy is that we estimatethe exposure time by calculating the ratio of mean values ofthe sequential images
The steps of generating HDR image are as follows
(1) Separate the corrected pathology image into H chan-nel stained byHaematoxylin and E channel stained byEosin based on the stain separation method
(2) Estimate the ratio of exposure time of adjacent imagesbased on their means of pixels for both correctedpathology images and H and E channel images Thenfit curve of pixel relationship based on least squaremethod
(3) Calculate camera response function based on Mit-sunaga algorithm [26] by replacing the pixel valuewith pixel relationship curve and then generate HDRpathological image
34 Detail Enhancement Based on GIF The enhancementmethod based on BF [27] can reduce the loss of image detailinformation but it has gradient reverse problem In this paperwe replace BFwithGIF becauseGIF can avoid image gradientreverse and reduce calculation cost In order to prove thesuperiority of GIF we design five groups of images withdifferent sizes And there are 10 color pathological imageswith the same size in each group Every image is processedwith BF and GIF respectively and the run time is recordedsuccessively Finally we calculate the average run time forevery group The comparison result is shown in Table 1
It is clear that GIF takes much less run time than BFwhenprocessing the same image Therefore in our image enhance-ment method we firstly transform the HDR pathology imageinto YCbCr color space (the type of sampling is 4 4 4) thendivide the luminance component into base layer and detaillayer by GIF rather than BF next compress the dynamic rangefor base layer using histogram mapping to reduce noise andenhance detail information meanwhile enhance the detaillayer by adaptive masking and finally combine the enhancedluminance component and color component together Thebasic procedure is shown in Figure 2
The equation of hierarchical processing for image basedon GIF is shown as follows
where 119865119866 is the GIF 119868119884 is the logarithm of luminance 119884 aftertransforming the HDR image into YCbCr color space 119868119861 isthe base layer of image including the grey information of tex-ture region in image and 119868119863 is the detail layer including high-frequency detail information such as edges (119909 119910) describethe coordinate of the pixel
The result of GIF is related to the radius 119903 of filter windowand the regular parameter (smooth factor) 120576 Figure 3 showshow these two parameters affect the hierarchical resultsfor pathology image Figure 3(a) is the original pathologicalimage and (b) is the Y channel image by transforming theoriginal image into YCbCr color space (c) (e) and (g) showthe base layer with 119903 = 6 120576 = 004 119903 = 10 120576 = 001and 119903 = 10 120576 = 004 respectively (d) (f) and (h) showthe detail layer of image corresponding to (c) (e) and (g)respectively It is seen that the detail layer with larger windowradius and regular parameter can obtain much more detailinformation
4 Experiments and Comparison
In this paper all the experiments are implemented usingMatlab R2014a development tool in the 64-bit Windows 7operating system (8-core CPU 340GHz 8 G memory) andour experimental data is actual clinical pathological imageswhich are RGB color images stained by Haematoxylin andEosin (H and E) The picture format is TIFF and spatialresolution is 1280 lowast 960 The data includes 123 liver tissueimages and 17 lung tissue images that is 140 images intotal Besides the parameters of the GIF are 119903 = 10 120576 =004 The enhancement result of our proposed method isshown in Figure 4 Figure 4(a) is the original pathologicalimage and Figure 4(b) is the reference image which has goodstain quality and is used in the stain normalization processFigure 4(c) is the normalized image after stain normalizationFigure 4(d) is the result after denoising and correcting byimproved bias field model Figures 4(e) and 4(f) show theE channel and H channel images after stain separation ofFigure 4(d) respectively Figure 4(g) is the HDR pathologicalimage generated with Figures 4(c) 4(e) and 4(f) displayedby TMO Figure 4(h) is the final detail enhanced HDRpathological image
It is observed that our proposed method can obtain agood image enhancement effect in Figure 4This method canimprove image luminance and contrast meanwhile it canpreserve the image detail structure well In order to verifythe effectiveness of our algorithm we compare our methodwith various Retinex-based methods and some HE-basedalgorithms namely Frankle-McCann Retinex [11] SSR [28]MSR [29] Dualistic Sub-Image Histogram Equalization
BioMed Research International 5
Transform into YCbCr space
Color component (CbCr)
Guided image filter
Base layer Detail layer
Histogram mapping Adaptive enhancement
Base layer in LDR Detail layer in LDR
Detail enhanced LDR luminance component
Merge
HDR pathological image
LDR image
Logarithm fetch on luminance component (Y)
Figure 2 The flowchart of detail enhancement based on GIF
(a) Original image (b) Y channel image (c) Base layer image (119903 = 6 120576 = 004) (d) Detail layer image (119903 = 6 120576 =004)
(e) Base layer image (119903 = 10 120576 =001)
(f) Detail layer image (119903 = 10 120576 =001)
(g) Base layer image (119903 = 10 120576 =004)
(h) Detail layer image (119903 = 10 120576 =004)
Figure 3 Illustration of layer separation on pathological image with different parameters
(DSIHE) [30] Minimum Mean Brightness Error Bi-Histo-gram Equalization (MMBEBHE) [31] Recursive Mean-Separate Histogram Equalization (RMSHE) [32] and Recur-sive Sub-Image Histogram Equalization (RSIHE) [33] Allthese methods are applied to our 140 pathological images
successively andwefinally get 140 groups of enhanced imagesThere are 9 images in every group one original image and 8result images enhanced by 8 image enhancement algorithmsFigure 5 demonstrates one of the groups where Figure 5(a) isthe original image Figure 5(b)ndashFigure 5(h) are the enhanced
6 BioMed Research International
(a) Original image (b) Reference image for stain normalization
(c) Normalized image (d) Corrected image by improved bias field model
(e) E channel image (f) H channel image
(g) HDR image after tone mapping (h) Detail enhanced HDR image
Figure 4 The result of our proposed HDR pathological image enhancement method
image by the above comparison algorithms respectively andFigure 5(i) is our result
We can see from Figure 5 that all the above methodscan enhance the pathological images to some extent But the
differences between the results are a little big To evaluatecorrectly the enhancement results of different algorithmsthis paper analyze the corresponding images from both thesubjective and objective aspects
BioMed Research International 7
(a) Original image (b) Frankle-McCann Retinex (c) SSR
(d) MSR (e) DSIHE (f) MMBEBHE
(g) RMSHE (h) RSIHE (i) Proposed
Figure 5 The comparison of our method with other different image enhancement methods
Table 2 The number of best enhanced image selected by pathologists for different algorithms
41 Subjective Evaluation Pathological image is the goldstandard in disease diagnosis In order to verify the prac-ticability of our algorithm we extract 50 groups of imagesrandomly and invite three pathologists to choose one bestimage from every group The selected image should be theoptimal in visual quality and the most helpful to clinicalapplication from the perspective of a pathologist The finalresult is shown in Table 2
As Table 2 shows the results of HE-based algorithmsare basically not recognized by pathologists because there isserious distortion problem in the enhanced image of thosealgorithms Most of the detail information in those images islost and the cell regions cannot be distinguished Although itis possible to obtain the optimal enhancement effect for everyone of the other algorithms our proposed method is betterstatistically having the maximum amount of the selected best
8 BioMed Research International
Table 3 Data comparison of our method with other different image enhancement methods
Algorithms EvaluationsPSNR (dB) SD Mean EME Entropy (bit) Run time (second)
image So from a pathologistrsquos viewpoint our pathologicalimage enhancement method can improve the image qualitya lot
42 Objective Evaluation On the other hand we also analyzethe experiment results quantitatively in five well-known met-rics namely peak signal noise ratio (PSNR) [34] standarddeviation (SD) mean measure of enhancement (EME) [35]and information entropy [34]
PSNR is widely used to evaluate the quality of imageThehigher PSNR value denotes that the image could suppressnoise better The calculation equation of PSNR is as follows
and where 119883 and 119884 are input image and output imagerespectively119873 is the total number of image pixels and119871 is thedynamic range of pixel values The unit of PSNR is decibels(dB)
Themean value is used to evaluate the average luminanceof image The higher mean value represents that the image isbrighter SD is also a popular metric in image enhancementand is used to estimate the contrast of image The imagecontrast is grater if the SD value is higher For mean and SDwe compute the average value of three channels of the colorpathological image in our experiments
EME is one well-known blind-reference image qualityassessment (IQA)metric It gives a quality score to each imagebased on the image contrast The larger EME value representsthe more detail information and more obvious variation inlocal region The definition of EME is as follows
EME11989611198962 =1
1198961 11989621198962
sum119897=1
1198961
sum119896=1
20 log 119868120596max119896119897
119868120596min119896119897 (13)
where the test image is divided into 1198961 times 1198962 small blocksand 119868120596max119896119897 and 119868120596min119896119897 represent the maximum andminimumvalues of pixel respectively in block 120596119896119897
Information entropy is an important metric to measurethe content of image And the higher value indicates an imagewith richer details The equation of entropy is as follows
119867 = minus119872
sum119896=1
119901119896log2119901119896 (14)
where119872 is the gray levels of image and 119901119896 is the probabilityof gray level 119896 in the whole image
All the objective evaluation result data is shown in Table 3and all the data is an average value of our 140 pathologicalimages We can see from Table 3 that all these algorithmscould enhance image to some extent The SD values of HE-based methods are much higher than other methods Buttheir contribution is not outstanding when taking the seriousdistortion problem of those enhanced images into accountIn addition the PSNR value of our algorithm is not as goodas the Retinex-based algorithms but we obtain the bestresults in all the other metrics of mean EME and entropyAnd considering that we just use the original image andthe enhanced image to replace the noise-free image and testimage in the definition of PSNR the contribution of PSNR islimited So in conclusion our method is superior to others
Moreover in order to indicate the complexity of ouralgorithm we compare the run times of all the algorithmsrunning in the same platform and using the same imageas described above The final average result is shown in thelast column of Table 3 It is obvious that our algorithm ismuch faster than Retinex-based algorithms Although therun time of HE-based algorithm is the least the time cost ofour algorithm is worthy in view of the quality of enhancedimages
43 Cell Segmentation We also use a simple automaticcell segmentation method based on morphology to ver-ify that our image enhancement algorithm can improvesubsequent image segmentation and other image analysisprocesses In this cell segmentation method firstly transformthe original RGB image into gray-scale image and thenconduct image opening operation image reconstructionimage binarization image erosion and dilation operationsand image denoising [36] We select 20 groups of enhanced
BioMed Research International 9
(a) Original image (b) Ground truth (c) Frankle-McCann (d) SSR (e) MSR
images in which the cell regions are relatively obvious to dotest Figure 6 demonstrates one group segmentation resultwhere Figure 6(a) is the original RGB pathological imageFigure 6(b) is the ground truth segmented manually underthe guidance of pathologists and Figure 6(c)ndashFigure 6(j)are obtained from segmenting the corresponding enhancedimages
It is clear that the cell segmentation result of ouralgorithm is more close to the ground truth In order tocompare the segmentation results quantitatively we adopttwo standard segmentation metrics to evaluate namely dicecoefficient [37] and Jaccard index [38] The definitions ofthem are as follows
DC (SRTR) = 2 timesNum (pixelSR cap pixelTR)Num (pixelSR) +Num (pixelTR)
JI (SRTR) = Num (pixelSR cap pixelTR)Num (pixelSR cup pixelTR)
(15)
where SR is the segmentation region TR is the true regionof the target and Num(pixel) is the number of relevantpixel points The average results of performance metricsare displayed in Table 4 We can see that the result ofour proposed method is the best Our HDR pathologicalimage enhancement method is superior to the comparisonalgorithms as we can improve the image segmentation andpathological analysis better
In sum we proposed that HDR pathological imageenhancement method obtains a better result according to
both pathologistsrsquo subjective evaluation and quantitativeanalysis in data and the cell segmentationmethod also provesthe better performance of our method as the quality anddetail of original image are both improved
5 Conclusions
This paper proposes new HDR pathological image enhance-ment methods based on GIF and improved bias fieldcorrection model First stain normalization and waveletdenoising operations are used in image preprocessing Andthe improved bias field model is introduced to correct theintensity inhomogeneity and detail discontinuity of imageThen the HDR pathological image is generated using LDRimage and H and E channel images Next the 119884 componentof HDR image is separated into base layer and detail layerby GIF and the two layers are enhanced separately Finallythe fine enhanced image is acquired after combining the 119884component and the color components To verify the effective-ness of the proposed method we perform the enhancementexperiments using 140 pathological images The experimentresults and comparisons with related work demonstrate thatour proposed method improves the image quality in terms ofhuman vision PSNR SD mean EME information entropyand cell segmentation
Competing Interests
The authors declare that there were no competing interestsregarding the publication of this article
10 BioMed Research International
Acknowledgments
This work was supported by the National Natural ScienceFoundation of China (no 61472073)
References
[1] SM Pizer E P Amburn J D Austin et al ldquoAdaptive histogramequalization and its variationsrdquo Computer Vision Graphics ampImage Processing vol 39 no 3 pp 355ndash368 1987
[2] V Vani and K V M Prashanth ldquoColor image enhancementtechniques in Wireless Capsule Endoscopyrdquo in Proceedings ofthe IEEE International Conference on Trends in AutomationCommunications and Computing Technology (I-TACT rsquo15) vol1 pp 1ndash6 Bangalore India December 2015
[3] H Cao L Tian J Liu H Wang and S Feng ldquoColor imageenhancement using power-constraint histogram equalizationfor AMOLEDrdquo in Proceedings of the IEEE 11th InternationalConference on ASIC (ASICON rsquo15) pp 1ndash4 IEEE ChengduChina November 2015
[4] N M Kwok G Fang and H Y Shi ldquoColor enhancementfor images from digital camera using a transformation-freeapproachrdquo in Proceedings of the 9th International Conferenceon Sensing Technology (ICST rsquo15) pp 168ndash172 IEEE AucklandNew Zealand December 2015
[5] S D Nikam and R U Yawale ldquoColor image enhancementusing daubechies wavelet transform and HIS color modelrdquoin Proceedings of the International Conference on IndustrialInstrumentation and Control (ICIC rsquo15) pp 1323ndash1327 IEEEPune India May 2015
[6] L G Villanueva G M Callico F Tobajas et al ldquoMedicaldiagnosis improvement through image quality enhancementbased on super-resolutionrdquo in Proceedings of the 13th EuromicroConference onDigital SystemDesign ArchitecturesMethods andTools (DSD rsquo10) pp 259ndash262 IEEE Lille France September2010
[7] W Sun F Li and Q Zhang ldquoThe applications of improvedretinex algorithm for X-ray medical image enhancementrdquo inProceedings of the International Conference on Computer Scienceand Service System (CSSS rsquo12) pp 1655ndash1658 IEEE NanjingChina August 2012
[8] G Zhang D Sun P Yan H Zhao and Z Li ldquoA LDCT imagecontrast enhancement algorithm based on single-scale retinextheoryrdquo in Proceedings of the International Conference on Com-putational Intelligence for Modelling Control amp Automation pp1282ndash1287 IEEE Computer Society Vienna Austria December2008
[9] S Setty N K Srinath and M C Hanumantharaju ldquoDevel-opment of multiscale retinex algorithm for medical imageenhancement based on multi-rate samplingrdquo in Proceedingsof the International Conference on Signal Processing ImageProcessing amp Pattern Recognition pp 145ndash150 2013
[10] J Mccann ldquoLessons learned from mondrians applied to realimages and color gamutsrdquo in Proceedings of the Color andImaging Conference vol 8 pp 1ndash8 1999
[11] B V Funt F Ciurea and J J McCann ldquoRetinex in Matlabrdquo inProceedings of the Color and Imaging Conference pp 112ndash121Scottsdale Ariz USA November 2000
[12] K Kim J Bae and J Kim ldquoNatural hdr image tone mappingbased on retinexrdquo IEEE Transactions on Consumer Electronicsvol 57 no 4 pp 1807ndash1814 2011
[13] J Kuang G M Johnson and M D Fairchild ldquoiCAM06 arefined image appearance model for HDR image renderingrdquoJournal of Visual Communication amp Image Representation vol18 no 5 pp 406ndash414 2007
[14] M-L Song H-Q Wang C Chen X-Q Ye and W-K GuldquoTone mapping for high dynamic range image using a proba-bilistic modelrdquo Journal of Software vol 20 no 3 pp 734ndash7432010
[15] F Branchitta M Diani G Corsini and M Romagnoli ldquoNewtechnique for the visualization of high dynamic range infraredimagesrdquo Optical Engineering vol 48 no 9 Article ID 0964012009
[16] C Zuo ldquoDisplay and detail enhancement for high-dynamic-range infrared imagesrdquo Optical Engineering vol 50 no 12Article ID 127401 pp 895ndash900 2011
[17] K He J Sun and X Tang ldquoGuided image filteringrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol35 no 6 pp 1397ndash1409 2013
[18] N Liu and D Zhao ldquoDetail enhancement for high-dynamic-range infrared images based on guided image filterrdquo InfraredPhysics amp Technology vol 67 pp 138ndash147 2014
[19] C Li R Huang Z Ding et al ldquoA level set method for imagesegmentation in the presence of intensity inhomogeneities withapplication toMRIrdquo IEEE Transactions on Image Processing vol20 no 7 pp 2007ndash2016 2011
[20] C Li J C Gore and C Davatzikos ldquoMultiplicative intrinsiccomponent optimization (MICO) for MRI bias field estimationand tissue segmentationrdquoMagnetic Resonance Imaging vol 32no 7 pp 913ndash923 2014
[21] A Vahadane T Peng A Sethi et al ldquoStructure-preservingcolor normalization and sparse stain separation for histologicalimagesrdquo IEEE Transactions on Medical Imaging vol 35 no 8pp 1962ndash1971 2016
[22] E Reinhard M Ashikhmin B Gooch and P Shirley ldquoColortransfer between imagesrdquo IEEE Computer Graphics amp Applica-tions vol 21 no 5 pp 34ndash41 2001
[23] D L Donoho ldquoDe-noising by soft-thresholdingrdquo IEEE Trans-actions on Information Theory vol 41 no 3 pp 613ndash627 1995
[24] E Zhang H Yang and M Xu ldquoA novel tone mappingmethod for high dynamic range image by incorporating edge-preserving filter into method based on retinexrdquo Applied Mathe-matics amp Information Sciences vol 9 no 1 pp 411ndash417 2015
[25] M Macenko M Niethammer J S Marron et al ldquoA methodfor normalizing histology slides for quantitative analysisrdquo inProceedings of the IEEE International Conference on Symposiumon Biomedical Imaging From Nano To Macro pp 1107ndash1110IEEE Press 2009
[26] T Mitsunaga and S K Nayar ldquoRadiometric self calibrationrdquoin Proceedings of the IEEE Computer Society Conference onComputer Vision and Pattern Recognition vol 1 p 1374 FortCollins Colo USA June 1999
[27] F Durand and J Dorsey ldquoFast bilateral filtering for the displayof high-dynamic-range imagesrdquoACMTransactions onGraphicsvol 21 no 3 pp 257ndash266 2002
[28] D J Jobson Z-U Rahman andG AWoodell ldquoProperties andperformance of a centersurround retinexrdquo IEEE Transactionson Image Processing vol 6 no 3 pp 451ndash462 1997
[29] D J Jobson Z-U Rahman and G A Woodell ldquoA multiscaleretinex for bridging the gap between color images and thehuman observation of scenesrdquo IEEE Transactions on ImageProcessing vol 6 no 7 pp 965ndash976 1997
BioMed Research International 11
[30] YWangQ Chen andB Zhang ldquoImage enhancement based onequal area dualistic sub-image histogram equalizationmethodrdquoIEEE Transactions on Consumer Electronics vol 45 no 1 pp68ndash75 1999
[31] S-D Chen and A R Ramli ldquoMinimum mean brightnesserror bi-histogram equalization in contrast enhancementrdquo IEEETransactions on Consumer Electronics vol 49 no 4 pp 1310ndash1319 2003
[32] S-D Chen and A R Ramli ldquoContrast enhancement usingrecursive mean-separate histogram equalization for scalablebrightness preservationrdquo IEEE Transactions on Consumer Elec-tronics vol 49 no 4 pp 1301ndash1309 2003
[33] K S Sim C P Tso and Y Y Tan ldquoRecursive sub-imagehistogram equalization applied to gray scale imagesrdquo PatternRecognition Letters vol 28 no 10 pp 1209ndash1221 2007
[34] V L Jaya and R Gopikakumari ldquoIEM a new image enhance-ment metric for contrast and sharpness measurementsrdquo Inter-national Journal of Computer Applications vol 79 no 9 pp 1ndash92013
[35] S S Agaian K Panetta and A M Grigoryan ldquoTransform-based image enhancement algorithms with performance mea-surerdquo IEEE Transactions on Image Processing vol 10 no 3 pp367ndash382 2001
[36] M Braiki A Benzinou K Nasreddine S Labidi and NHymery ldquoSegmentation of dendritic cells from microscopicimages using mathematical morphologyrdquo in Proceedings ofthe 2nd International Conference on Advanced Technologies forSignal and Image Processing (ATSIP rsquo16) pp 282ndash287 MonastirTunisia March 2016
[37] L R Dice ldquoMeasures of the amount of ecologic associationbetween speciesrdquo Ecology vol 26 no 3 pp 297ndash302 1945
[38] P Jaccard ldquoThe distribution of the flora in the alpine zonerdquoNewPhytologist vol 11 no 2 pp 37ndash50 1912
We usually need many image sequences with differentexposure of the same scene to generate HDR image But inthis paper we generate HDR image sequences based on stainseparationmethod [25] In this case we cannot determine theexposure time Here one sample strategy is that we estimatethe exposure time by calculating the ratio of mean values ofthe sequential images
The steps of generating HDR image are as follows
(1) Separate the corrected pathology image into H chan-nel stained byHaematoxylin and E channel stained byEosin based on the stain separation method
(2) Estimate the ratio of exposure time of adjacent imagesbased on their means of pixels for both correctedpathology images and H and E channel images Thenfit curve of pixel relationship based on least squaremethod
(3) Calculate camera response function based on Mit-sunaga algorithm [26] by replacing the pixel valuewith pixel relationship curve and then generate HDRpathological image
34 Detail Enhancement Based on GIF The enhancementmethod based on BF [27] can reduce the loss of image detailinformation but it has gradient reverse problem In this paperwe replace BFwithGIF becauseGIF can avoid image gradientreverse and reduce calculation cost In order to prove thesuperiority of GIF we design five groups of images withdifferent sizes And there are 10 color pathological imageswith the same size in each group Every image is processedwith BF and GIF respectively and the run time is recordedsuccessively Finally we calculate the average run time forevery group The comparison result is shown in Table 1
It is clear that GIF takes much less run time than BFwhenprocessing the same image Therefore in our image enhance-ment method we firstly transform the HDR pathology imageinto YCbCr color space (the type of sampling is 4 4 4) thendivide the luminance component into base layer and detaillayer by GIF rather than BF next compress the dynamic rangefor base layer using histogram mapping to reduce noise andenhance detail information meanwhile enhance the detaillayer by adaptive masking and finally combine the enhancedluminance component and color component together Thebasic procedure is shown in Figure 2
The equation of hierarchical processing for image basedon GIF is shown as follows
where 119865119866 is the GIF 119868119884 is the logarithm of luminance 119884 aftertransforming the HDR image into YCbCr color space 119868119861 isthe base layer of image including the grey information of tex-ture region in image and 119868119863 is the detail layer including high-frequency detail information such as edges (119909 119910) describethe coordinate of the pixel
The result of GIF is related to the radius 119903 of filter windowand the regular parameter (smooth factor) 120576 Figure 3 showshow these two parameters affect the hierarchical resultsfor pathology image Figure 3(a) is the original pathologicalimage and (b) is the Y channel image by transforming theoriginal image into YCbCr color space (c) (e) and (g) showthe base layer with 119903 = 6 120576 = 004 119903 = 10 120576 = 001and 119903 = 10 120576 = 004 respectively (d) (f) and (h) showthe detail layer of image corresponding to (c) (e) and (g)respectively It is seen that the detail layer with larger windowradius and regular parameter can obtain much more detailinformation
4 Experiments and Comparison
In this paper all the experiments are implemented usingMatlab R2014a development tool in the 64-bit Windows 7operating system (8-core CPU 340GHz 8 G memory) andour experimental data is actual clinical pathological imageswhich are RGB color images stained by Haematoxylin andEosin (H and E) The picture format is TIFF and spatialresolution is 1280 lowast 960 The data includes 123 liver tissueimages and 17 lung tissue images that is 140 images intotal Besides the parameters of the GIF are 119903 = 10 120576 =004 The enhancement result of our proposed method isshown in Figure 4 Figure 4(a) is the original pathologicalimage and Figure 4(b) is the reference image which has goodstain quality and is used in the stain normalization processFigure 4(c) is the normalized image after stain normalizationFigure 4(d) is the result after denoising and correcting byimproved bias field model Figures 4(e) and 4(f) show theE channel and H channel images after stain separation ofFigure 4(d) respectively Figure 4(g) is the HDR pathologicalimage generated with Figures 4(c) 4(e) and 4(f) displayedby TMO Figure 4(h) is the final detail enhanced HDRpathological image
It is observed that our proposed method can obtain agood image enhancement effect in Figure 4This method canimprove image luminance and contrast meanwhile it canpreserve the image detail structure well In order to verifythe effectiveness of our algorithm we compare our methodwith various Retinex-based methods and some HE-basedalgorithms namely Frankle-McCann Retinex [11] SSR [28]MSR [29] Dualistic Sub-Image Histogram Equalization
BioMed Research International 5
Transform into YCbCr space
Color component (CbCr)
Guided image filter
Base layer Detail layer
Histogram mapping Adaptive enhancement
Base layer in LDR Detail layer in LDR
Detail enhanced LDR luminance component
Merge
HDR pathological image
LDR image
Logarithm fetch on luminance component (Y)
Figure 2 The flowchart of detail enhancement based on GIF
(a) Original image (b) Y channel image (c) Base layer image (119903 = 6 120576 = 004) (d) Detail layer image (119903 = 6 120576 =004)
(e) Base layer image (119903 = 10 120576 =001)
(f) Detail layer image (119903 = 10 120576 =001)
(g) Base layer image (119903 = 10 120576 =004)
(h) Detail layer image (119903 = 10 120576 =004)
Figure 3 Illustration of layer separation on pathological image with different parameters
(DSIHE) [30] Minimum Mean Brightness Error Bi-Histo-gram Equalization (MMBEBHE) [31] Recursive Mean-Separate Histogram Equalization (RMSHE) [32] and Recur-sive Sub-Image Histogram Equalization (RSIHE) [33] Allthese methods are applied to our 140 pathological images
successively andwefinally get 140 groups of enhanced imagesThere are 9 images in every group one original image and 8result images enhanced by 8 image enhancement algorithmsFigure 5 demonstrates one of the groups where Figure 5(a) isthe original image Figure 5(b)ndashFigure 5(h) are the enhanced
6 BioMed Research International
(a) Original image (b) Reference image for stain normalization
(c) Normalized image (d) Corrected image by improved bias field model
(e) E channel image (f) H channel image
(g) HDR image after tone mapping (h) Detail enhanced HDR image
Figure 4 The result of our proposed HDR pathological image enhancement method
image by the above comparison algorithms respectively andFigure 5(i) is our result
We can see from Figure 5 that all the above methodscan enhance the pathological images to some extent But the
differences between the results are a little big To evaluatecorrectly the enhancement results of different algorithmsthis paper analyze the corresponding images from both thesubjective and objective aspects
BioMed Research International 7
(a) Original image (b) Frankle-McCann Retinex (c) SSR
(d) MSR (e) DSIHE (f) MMBEBHE
(g) RMSHE (h) RSIHE (i) Proposed
Figure 5 The comparison of our method with other different image enhancement methods
Table 2 The number of best enhanced image selected by pathologists for different algorithms
41 Subjective Evaluation Pathological image is the goldstandard in disease diagnosis In order to verify the prac-ticability of our algorithm we extract 50 groups of imagesrandomly and invite three pathologists to choose one bestimage from every group The selected image should be theoptimal in visual quality and the most helpful to clinicalapplication from the perspective of a pathologist The finalresult is shown in Table 2
As Table 2 shows the results of HE-based algorithmsare basically not recognized by pathologists because there isserious distortion problem in the enhanced image of thosealgorithms Most of the detail information in those images islost and the cell regions cannot be distinguished Although itis possible to obtain the optimal enhancement effect for everyone of the other algorithms our proposed method is betterstatistically having the maximum amount of the selected best
8 BioMed Research International
Table 3 Data comparison of our method with other different image enhancement methods
Algorithms EvaluationsPSNR (dB) SD Mean EME Entropy (bit) Run time (second)
image So from a pathologistrsquos viewpoint our pathologicalimage enhancement method can improve the image qualitya lot
42 Objective Evaluation On the other hand we also analyzethe experiment results quantitatively in five well-known met-rics namely peak signal noise ratio (PSNR) [34] standarddeviation (SD) mean measure of enhancement (EME) [35]and information entropy [34]
PSNR is widely used to evaluate the quality of imageThehigher PSNR value denotes that the image could suppressnoise better The calculation equation of PSNR is as follows
and where 119883 and 119884 are input image and output imagerespectively119873 is the total number of image pixels and119871 is thedynamic range of pixel values The unit of PSNR is decibels(dB)
Themean value is used to evaluate the average luminanceof image The higher mean value represents that the image isbrighter SD is also a popular metric in image enhancementand is used to estimate the contrast of image The imagecontrast is grater if the SD value is higher For mean and SDwe compute the average value of three channels of the colorpathological image in our experiments
EME is one well-known blind-reference image qualityassessment (IQA)metric It gives a quality score to each imagebased on the image contrast The larger EME value representsthe more detail information and more obvious variation inlocal region The definition of EME is as follows
EME11989611198962 =1
1198961 11989621198962
sum119897=1
1198961
sum119896=1
20 log 119868120596max119896119897
119868120596min119896119897 (13)
where the test image is divided into 1198961 times 1198962 small blocksand 119868120596max119896119897 and 119868120596min119896119897 represent the maximum andminimumvalues of pixel respectively in block 120596119896119897
Information entropy is an important metric to measurethe content of image And the higher value indicates an imagewith richer details The equation of entropy is as follows
119867 = minus119872
sum119896=1
119901119896log2119901119896 (14)
where119872 is the gray levels of image and 119901119896 is the probabilityof gray level 119896 in the whole image
All the objective evaluation result data is shown in Table 3and all the data is an average value of our 140 pathologicalimages We can see from Table 3 that all these algorithmscould enhance image to some extent The SD values of HE-based methods are much higher than other methods Buttheir contribution is not outstanding when taking the seriousdistortion problem of those enhanced images into accountIn addition the PSNR value of our algorithm is not as goodas the Retinex-based algorithms but we obtain the bestresults in all the other metrics of mean EME and entropyAnd considering that we just use the original image andthe enhanced image to replace the noise-free image and testimage in the definition of PSNR the contribution of PSNR islimited So in conclusion our method is superior to others
Moreover in order to indicate the complexity of ouralgorithm we compare the run times of all the algorithmsrunning in the same platform and using the same imageas described above The final average result is shown in thelast column of Table 3 It is obvious that our algorithm ismuch faster than Retinex-based algorithms Although therun time of HE-based algorithm is the least the time cost ofour algorithm is worthy in view of the quality of enhancedimages
43 Cell Segmentation We also use a simple automaticcell segmentation method based on morphology to ver-ify that our image enhancement algorithm can improvesubsequent image segmentation and other image analysisprocesses In this cell segmentation method firstly transformthe original RGB image into gray-scale image and thenconduct image opening operation image reconstructionimage binarization image erosion and dilation operationsand image denoising [36] We select 20 groups of enhanced
BioMed Research International 9
(a) Original image (b) Ground truth (c) Frankle-McCann (d) SSR (e) MSR
images in which the cell regions are relatively obvious to dotest Figure 6 demonstrates one group segmentation resultwhere Figure 6(a) is the original RGB pathological imageFigure 6(b) is the ground truth segmented manually underthe guidance of pathologists and Figure 6(c)ndashFigure 6(j)are obtained from segmenting the corresponding enhancedimages
It is clear that the cell segmentation result of ouralgorithm is more close to the ground truth In order tocompare the segmentation results quantitatively we adopttwo standard segmentation metrics to evaluate namely dicecoefficient [37] and Jaccard index [38] The definitions ofthem are as follows
DC (SRTR) = 2 timesNum (pixelSR cap pixelTR)Num (pixelSR) +Num (pixelTR)
JI (SRTR) = Num (pixelSR cap pixelTR)Num (pixelSR cup pixelTR)
(15)
where SR is the segmentation region TR is the true regionof the target and Num(pixel) is the number of relevantpixel points The average results of performance metricsare displayed in Table 4 We can see that the result ofour proposed method is the best Our HDR pathologicalimage enhancement method is superior to the comparisonalgorithms as we can improve the image segmentation andpathological analysis better
In sum we proposed that HDR pathological imageenhancement method obtains a better result according to
both pathologistsrsquo subjective evaluation and quantitativeanalysis in data and the cell segmentationmethod also provesthe better performance of our method as the quality anddetail of original image are both improved
5 Conclusions
This paper proposes new HDR pathological image enhance-ment methods based on GIF and improved bias fieldcorrection model First stain normalization and waveletdenoising operations are used in image preprocessing Andthe improved bias field model is introduced to correct theintensity inhomogeneity and detail discontinuity of imageThen the HDR pathological image is generated using LDRimage and H and E channel images Next the 119884 componentof HDR image is separated into base layer and detail layerby GIF and the two layers are enhanced separately Finallythe fine enhanced image is acquired after combining the 119884component and the color components To verify the effective-ness of the proposed method we perform the enhancementexperiments using 140 pathological images The experimentresults and comparisons with related work demonstrate thatour proposed method improves the image quality in terms ofhuman vision PSNR SD mean EME information entropyand cell segmentation
Competing Interests
The authors declare that there were no competing interestsregarding the publication of this article
10 BioMed Research International
Acknowledgments
This work was supported by the National Natural ScienceFoundation of China (no 61472073)
References
[1] SM Pizer E P Amburn J D Austin et al ldquoAdaptive histogramequalization and its variationsrdquo Computer Vision Graphics ampImage Processing vol 39 no 3 pp 355ndash368 1987
[2] V Vani and K V M Prashanth ldquoColor image enhancementtechniques in Wireless Capsule Endoscopyrdquo in Proceedings ofthe IEEE International Conference on Trends in AutomationCommunications and Computing Technology (I-TACT rsquo15) vol1 pp 1ndash6 Bangalore India December 2015
[3] H Cao L Tian J Liu H Wang and S Feng ldquoColor imageenhancement using power-constraint histogram equalizationfor AMOLEDrdquo in Proceedings of the IEEE 11th InternationalConference on ASIC (ASICON rsquo15) pp 1ndash4 IEEE ChengduChina November 2015
[4] N M Kwok G Fang and H Y Shi ldquoColor enhancementfor images from digital camera using a transformation-freeapproachrdquo in Proceedings of the 9th International Conferenceon Sensing Technology (ICST rsquo15) pp 168ndash172 IEEE AucklandNew Zealand December 2015
[5] S D Nikam and R U Yawale ldquoColor image enhancementusing daubechies wavelet transform and HIS color modelrdquoin Proceedings of the International Conference on IndustrialInstrumentation and Control (ICIC rsquo15) pp 1323ndash1327 IEEEPune India May 2015
[6] L G Villanueva G M Callico F Tobajas et al ldquoMedicaldiagnosis improvement through image quality enhancementbased on super-resolutionrdquo in Proceedings of the 13th EuromicroConference onDigital SystemDesign ArchitecturesMethods andTools (DSD rsquo10) pp 259ndash262 IEEE Lille France September2010
[7] W Sun F Li and Q Zhang ldquoThe applications of improvedretinex algorithm for X-ray medical image enhancementrdquo inProceedings of the International Conference on Computer Scienceand Service System (CSSS rsquo12) pp 1655ndash1658 IEEE NanjingChina August 2012
[8] G Zhang D Sun P Yan H Zhao and Z Li ldquoA LDCT imagecontrast enhancement algorithm based on single-scale retinextheoryrdquo in Proceedings of the International Conference on Com-putational Intelligence for Modelling Control amp Automation pp1282ndash1287 IEEE Computer Society Vienna Austria December2008
[9] S Setty N K Srinath and M C Hanumantharaju ldquoDevel-opment of multiscale retinex algorithm for medical imageenhancement based on multi-rate samplingrdquo in Proceedingsof the International Conference on Signal Processing ImageProcessing amp Pattern Recognition pp 145ndash150 2013
[10] J Mccann ldquoLessons learned from mondrians applied to realimages and color gamutsrdquo in Proceedings of the Color andImaging Conference vol 8 pp 1ndash8 1999
[11] B V Funt F Ciurea and J J McCann ldquoRetinex in Matlabrdquo inProceedings of the Color and Imaging Conference pp 112ndash121Scottsdale Ariz USA November 2000
[12] K Kim J Bae and J Kim ldquoNatural hdr image tone mappingbased on retinexrdquo IEEE Transactions on Consumer Electronicsvol 57 no 4 pp 1807ndash1814 2011
[13] J Kuang G M Johnson and M D Fairchild ldquoiCAM06 arefined image appearance model for HDR image renderingrdquoJournal of Visual Communication amp Image Representation vol18 no 5 pp 406ndash414 2007
[14] M-L Song H-Q Wang C Chen X-Q Ye and W-K GuldquoTone mapping for high dynamic range image using a proba-bilistic modelrdquo Journal of Software vol 20 no 3 pp 734ndash7432010
[15] F Branchitta M Diani G Corsini and M Romagnoli ldquoNewtechnique for the visualization of high dynamic range infraredimagesrdquo Optical Engineering vol 48 no 9 Article ID 0964012009
[16] C Zuo ldquoDisplay and detail enhancement for high-dynamic-range infrared imagesrdquo Optical Engineering vol 50 no 12Article ID 127401 pp 895ndash900 2011
[17] K He J Sun and X Tang ldquoGuided image filteringrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol35 no 6 pp 1397ndash1409 2013
[18] N Liu and D Zhao ldquoDetail enhancement for high-dynamic-range infrared images based on guided image filterrdquo InfraredPhysics amp Technology vol 67 pp 138ndash147 2014
[19] C Li R Huang Z Ding et al ldquoA level set method for imagesegmentation in the presence of intensity inhomogeneities withapplication toMRIrdquo IEEE Transactions on Image Processing vol20 no 7 pp 2007ndash2016 2011
[20] C Li J C Gore and C Davatzikos ldquoMultiplicative intrinsiccomponent optimization (MICO) for MRI bias field estimationand tissue segmentationrdquoMagnetic Resonance Imaging vol 32no 7 pp 913ndash923 2014
[21] A Vahadane T Peng A Sethi et al ldquoStructure-preservingcolor normalization and sparse stain separation for histologicalimagesrdquo IEEE Transactions on Medical Imaging vol 35 no 8pp 1962ndash1971 2016
[22] E Reinhard M Ashikhmin B Gooch and P Shirley ldquoColortransfer between imagesrdquo IEEE Computer Graphics amp Applica-tions vol 21 no 5 pp 34ndash41 2001
[23] D L Donoho ldquoDe-noising by soft-thresholdingrdquo IEEE Trans-actions on Information Theory vol 41 no 3 pp 613ndash627 1995
[24] E Zhang H Yang and M Xu ldquoA novel tone mappingmethod for high dynamic range image by incorporating edge-preserving filter into method based on retinexrdquo Applied Mathe-matics amp Information Sciences vol 9 no 1 pp 411ndash417 2015
[25] M Macenko M Niethammer J S Marron et al ldquoA methodfor normalizing histology slides for quantitative analysisrdquo inProceedings of the IEEE International Conference on Symposiumon Biomedical Imaging From Nano To Macro pp 1107ndash1110IEEE Press 2009
[26] T Mitsunaga and S K Nayar ldquoRadiometric self calibrationrdquoin Proceedings of the IEEE Computer Society Conference onComputer Vision and Pattern Recognition vol 1 p 1374 FortCollins Colo USA June 1999
[27] F Durand and J Dorsey ldquoFast bilateral filtering for the displayof high-dynamic-range imagesrdquoACMTransactions onGraphicsvol 21 no 3 pp 257ndash266 2002
[28] D J Jobson Z-U Rahman andG AWoodell ldquoProperties andperformance of a centersurround retinexrdquo IEEE Transactionson Image Processing vol 6 no 3 pp 451ndash462 1997
[29] D J Jobson Z-U Rahman and G A Woodell ldquoA multiscaleretinex for bridging the gap between color images and thehuman observation of scenesrdquo IEEE Transactions on ImageProcessing vol 6 no 7 pp 965ndash976 1997
BioMed Research International 11
[30] YWangQ Chen andB Zhang ldquoImage enhancement based onequal area dualistic sub-image histogram equalizationmethodrdquoIEEE Transactions on Consumer Electronics vol 45 no 1 pp68ndash75 1999
[31] S-D Chen and A R Ramli ldquoMinimum mean brightnesserror bi-histogram equalization in contrast enhancementrdquo IEEETransactions on Consumer Electronics vol 49 no 4 pp 1310ndash1319 2003
[32] S-D Chen and A R Ramli ldquoContrast enhancement usingrecursive mean-separate histogram equalization for scalablebrightness preservationrdquo IEEE Transactions on Consumer Elec-tronics vol 49 no 4 pp 1301ndash1309 2003
[33] K S Sim C P Tso and Y Y Tan ldquoRecursive sub-imagehistogram equalization applied to gray scale imagesrdquo PatternRecognition Letters vol 28 no 10 pp 1209ndash1221 2007
[34] V L Jaya and R Gopikakumari ldquoIEM a new image enhance-ment metric for contrast and sharpness measurementsrdquo Inter-national Journal of Computer Applications vol 79 no 9 pp 1ndash92013
[35] S S Agaian K Panetta and A M Grigoryan ldquoTransform-based image enhancement algorithms with performance mea-surerdquo IEEE Transactions on Image Processing vol 10 no 3 pp367ndash382 2001
[36] M Braiki A Benzinou K Nasreddine S Labidi and NHymery ldquoSegmentation of dendritic cells from microscopicimages using mathematical morphologyrdquo in Proceedings ofthe 2nd International Conference on Advanced Technologies forSignal and Image Processing (ATSIP rsquo16) pp 282ndash287 MonastirTunisia March 2016
[37] L R Dice ldquoMeasures of the amount of ecologic associationbetween speciesrdquo Ecology vol 26 no 3 pp 297ndash302 1945
[38] P Jaccard ldquoThe distribution of the flora in the alpine zonerdquoNewPhytologist vol 11 no 2 pp 37ndash50 1912
Figure 2 The flowchart of detail enhancement based on GIF
(a) Original image (b) Y channel image (c) Base layer image (119903 = 6 120576 = 004) (d) Detail layer image (119903 = 6 120576 =004)
(e) Base layer image (119903 = 10 120576 =001)
(f) Detail layer image (119903 = 10 120576 =001)
(g) Base layer image (119903 = 10 120576 =004)
(h) Detail layer image (119903 = 10 120576 =004)
Figure 3 Illustration of layer separation on pathological image with different parameters
(DSIHE) [30] Minimum Mean Brightness Error Bi-Histo-gram Equalization (MMBEBHE) [31] Recursive Mean-Separate Histogram Equalization (RMSHE) [32] and Recur-sive Sub-Image Histogram Equalization (RSIHE) [33] Allthese methods are applied to our 140 pathological images
successively andwefinally get 140 groups of enhanced imagesThere are 9 images in every group one original image and 8result images enhanced by 8 image enhancement algorithmsFigure 5 demonstrates one of the groups where Figure 5(a) isthe original image Figure 5(b)ndashFigure 5(h) are the enhanced
6 BioMed Research International
(a) Original image (b) Reference image for stain normalization
(c) Normalized image (d) Corrected image by improved bias field model
(e) E channel image (f) H channel image
(g) HDR image after tone mapping (h) Detail enhanced HDR image
Figure 4 The result of our proposed HDR pathological image enhancement method
image by the above comparison algorithms respectively andFigure 5(i) is our result
We can see from Figure 5 that all the above methodscan enhance the pathological images to some extent But the
differences between the results are a little big To evaluatecorrectly the enhancement results of different algorithmsthis paper analyze the corresponding images from both thesubjective and objective aspects
BioMed Research International 7
(a) Original image (b) Frankle-McCann Retinex (c) SSR
(d) MSR (e) DSIHE (f) MMBEBHE
(g) RMSHE (h) RSIHE (i) Proposed
Figure 5 The comparison of our method with other different image enhancement methods
Table 2 The number of best enhanced image selected by pathologists for different algorithms
41 Subjective Evaluation Pathological image is the goldstandard in disease diagnosis In order to verify the prac-ticability of our algorithm we extract 50 groups of imagesrandomly and invite three pathologists to choose one bestimage from every group The selected image should be theoptimal in visual quality and the most helpful to clinicalapplication from the perspective of a pathologist The finalresult is shown in Table 2
As Table 2 shows the results of HE-based algorithmsare basically not recognized by pathologists because there isserious distortion problem in the enhanced image of thosealgorithms Most of the detail information in those images islost and the cell regions cannot be distinguished Although itis possible to obtain the optimal enhancement effect for everyone of the other algorithms our proposed method is betterstatistically having the maximum amount of the selected best
8 BioMed Research International
Table 3 Data comparison of our method with other different image enhancement methods
Algorithms EvaluationsPSNR (dB) SD Mean EME Entropy (bit) Run time (second)
image So from a pathologistrsquos viewpoint our pathologicalimage enhancement method can improve the image qualitya lot
42 Objective Evaluation On the other hand we also analyzethe experiment results quantitatively in five well-known met-rics namely peak signal noise ratio (PSNR) [34] standarddeviation (SD) mean measure of enhancement (EME) [35]and information entropy [34]
PSNR is widely used to evaluate the quality of imageThehigher PSNR value denotes that the image could suppressnoise better The calculation equation of PSNR is as follows
and where 119883 and 119884 are input image and output imagerespectively119873 is the total number of image pixels and119871 is thedynamic range of pixel values The unit of PSNR is decibels(dB)
Themean value is used to evaluate the average luminanceof image The higher mean value represents that the image isbrighter SD is also a popular metric in image enhancementand is used to estimate the contrast of image The imagecontrast is grater if the SD value is higher For mean and SDwe compute the average value of three channels of the colorpathological image in our experiments
EME is one well-known blind-reference image qualityassessment (IQA)metric It gives a quality score to each imagebased on the image contrast The larger EME value representsthe more detail information and more obvious variation inlocal region The definition of EME is as follows
EME11989611198962 =1
1198961 11989621198962
sum119897=1
1198961
sum119896=1
20 log 119868120596max119896119897
119868120596min119896119897 (13)
where the test image is divided into 1198961 times 1198962 small blocksand 119868120596max119896119897 and 119868120596min119896119897 represent the maximum andminimumvalues of pixel respectively in block 120596119896119897
Information entropy is an important metric to measurethe content of image And the higher value indicates an imagewith richer details The equation of entropy is as follows
119867 = minus119872
sum119896=1
119901119896log2119901119896 (14)
where119872 is the gray levels of image and 119901119896 is the probabilityof gray level 119896 in the whole image
All the objective evaluation result data is shown in Table 3and all the data is an average value of our 140 pathologicalimages We can see from Table 3 that all these algorithmscould enhance image to some extent The SD values of HE-based methods are much higher than other methods Buttheir contribution is not outstanding when taking the seriousdistortion problem of those enhanced images into accountIn addition the PSNR value of our algorithm is not as goodas the Retinex-based algorithms but we obtain the bestresults in all the other metrics of mean EME and entropyAnd considering that we just use the original image andthe enhanced image to replace the noise-free image and testimage in the definition of PSNR the contribution of PSNR islimited So in conclusion our method is superior to others
Moreover in order to indicate the complexity of ouralgorithm we compare the run times of all the algorithmsrunning in the same platform and using the same imageas described above The final average result is shown in thelast column of Table 3 It is obvious that our algorithm ismuch faster than Retinex-based algorithms Although therun time of HE-based algorithm is the least the time cost ofour algorithm is worthy in view of the quality of enhancedimages
43 Cell Segmentation We also use a simple automaticcell segmentation method based on morphology to ver-ify that our image enhancement algorithm can improvesubsequent image segmentation and other image analysisprocesses In this cell segmentation method firstly transformthe original RGB image into gray-scale image and thenconduct image opening operation image reconstructionimage binarization image erosion and dilation operationsand image denoising [36] We select 20 groups of enhanced
BioMed Research International 9
(a) Original image (b) Ground truth (c) Frankle-McCann (d) SSR (e) MSR
images in which the cell regions are relatively obvious to dotest Figure 6 demonstrates one group segmentation resultwhere Figure 6(a) is the original RGB pathological imageFigure 6(b) is the ground truth segmented manually underthe guidance of pathologists and Figure 6(c)ndashFigure 6(j)are obtained from segmenting the corresponding enhancedimages
It is clear that the cell segmentation result of ouralgorithm is more close to the ground truth In order tocompare the segmentation results quantitatively we adopttwo standard segmentation metrics to evaluate namely dicecoefficient [37] and Jaccard index [38] The definitions ofthem are as follows
DC (SRTR) = 2 timesNum (pixelSR cap pixelTR)Num (pixelSR) +Num (pixelTR)
JI (SRTR) = Num (pixelSR cap pixelTR)Num (pixelSR cup pixelTR)
(15)
where SR is the segmentation region TR is the true regionof the target and Num(pixel) is the number of relevantpixel points The average results of performance metricsare displayed in Table 4 We can see that the result ofour proposed method is the best Our HDR pathologicalimage enhancement method is superior to the comparisonalgorithms as we can improve the image segmentation andpathological analysis better
In sum we proposed that HDR pathological imageenhancement method obtains a better result according to
both pathologistsrsquo subjective evaluation and quantitativeanalysis in data and the cell segmentationmethod also provesthe better performance of our method as the quality anddetail of original image are both improved
5 Conclusions
This paper proposes new HDR pathological image enhance-ment methods based on GIF and improved bias fieldcorrection model First stain normalization and waveletdenoising operations are used in image preprocessing Andthe improved bias field model is introduced to correct theintensity inhomogeneity and detail discontinuity of imageThen the HDR pathological image is generated using LDRimage and H and E channel images Next the 119884 componentof HDR image is separated into base layer and detail layerby GIF and the two layers are enhanced separately Finallythe fine enhanced image is acquired after combining the 119884component and the color components To verify the effective-ness of the proposed method we perform the enhancementexperiments using 140 pathological images The experimentresults and comparisons with related work demonstrate thatour proposed method improves the image quality in terms ofhuman vision PSNR SD mean EME information entropyand cell segmentation
Competing Interests
The authors declare that there were no competing interestsregarding the publication of this article
10 BioMed Research International
Acknowledgments
This work was supported by the National Natural ScienceFoundation of China (no 61472073)
References
[1] SM Pizer E P Amburn J D Austin et al ldquoAdaptive histogramequalization and its variationsrdquo Computer Vision Graphics ampImage Processing vol 39 no 3 pp 355ndash368 1987
[2] V Vani and K V M Prashanth ldquoColor image enhancementtechniques in Wireless Capsule Endoscopyrdquo in Proceedings ofthe IEEE International Conference on Trends in AutomationCommunications and Computing Technology (I-TACT rsquo15) vol1 pp 1ndash6 Bangalore India December 2015
[3] H Cao L Tian J Liu H Wang and S Feng ldquoColor imageenhancement using power-constraint histogram equalizationfor AMOLEDrdquo in Proceedings of the IEEE 11th InternationalConference on ASIC (ASICON rsquo15) pp 1ndash4 IEEE ChengduChina November 2015
[4] N M Kwok G Fang and H Y Shi ldquoColor enhancementfor images from digital camera using a transformation-freeapproachrdquo in Proceedings of the 9th International Conferenceon Sensing Technology (ICST rsquo15) pp 168ndash172 IEEE AucklandNew Zealand December 2015
[5] S D Nikam and R U Yawale ldquoColor image enhancementusing daubechies wavelet transform and HIS color modelrdquoin Proceedings of the International Conference on IndustrialInstrumentation and Control (ICIC rsquo15) pp 1323ndash1327 IEEEPune India May 2015
[6] L G Villanueva G M Callico F Tobajas et al ldquoMedicaldiagnosis improvement through image quality enhancementbased on super-resolutionrdquo in Proceedings of the 13th EuromicroConference onDigital SystemDesign ArchitecturesMethods andTools (DSD rsquo10) pp 259ndash262 IEEE Lille France September2010
[7] W Sun F Li and Q Zhang ldquoThe applications of improvedretinex algorithm for X-ray medical image enhancementrdquo inProceedings of the International Conference on Computer Scienceand Service System (CSSS rsquo12) pp 1655ndash1658 IEEE NanjingChina August 2012
[8] G Zhang D Sun P Yan H Zhao and Z Li ldquoA LDCT imagecontrast enhancement algorithm based on single-scale retinextheoryrdquo in Proceedings of the International Conference on Com-putational Intelligence for Modelling Control amp Automation pp1282ndash1287 IEEE Computer Society Vienna Austria December2008
[9] S Setty N K Srinath and M C Hanumantharaju ldquoDevel-opment of multiscale retinex algorithm for medical imageenhancement based on multi-rate samplingrdquo in Proceedingsof the International Conference on Signal Processing ImageProcessing amp Pattern Recognition pp 145ndash150 2013
[10] J Mccann ldquoLessons learned from mondrians applied to realimages and color gamutsrdquo in Proceedings of the Color andImaging Conference vol 8 pp 1ndash8 1999
[11] B V Funt F Ciurea and J J McCann ldquoRetinex in Matlabrdquo inProceedings of the Color and Imaging Conference pp 112ndash121Scottsdale Ariz USA November 2000
[12] K Kim J Bae and J Kim ldquoNatural hdr image tone mappingbased on retinexrdquo IEEE Transactions on Consumer Electronicsvol 57 no 4 pp 1807ndash1814 2011
[13] J Kuang G M Johnson and M D Fairchild ldquoiCAM06 arefined image appearance model for HDR image renderingrdquoJournal of Visual Communication amp Image Representation vol18 no 5 pp 406ndash414 2007
[14] M-L Song H-Q Wang C Chen X-Q Ye and W-K GuldquoTone mapping for high dynamic range image using a proba-bilistic modelrdquo Journal of Software vol 20 no 3 pp 734ndash7432010
[15] F Branchitta M Diani G Corsini and M Romagnoli ldquoNewtechnique for the visualization of high dynamic range infraredimagesrdquo Optical Engineering vol 48 no 9 Article ID 0964012009
[16] C Zuo ldquoDisplay and detail enhancement for high-dynamic-range infrared imagesrdquo Optical Engineering vol 50 no 12Article ID 127401 pp 895ndash900 2011
[17] K He J Sun and X Tang ldquoGuided image filteringrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol35 no 6 pp 1397ndash1409 2013
[18] N Liu and D Zhao ldquoDetail enhancement for high-dynamic-range infrared images based on guided image filterrdquo InfraredPhysics amp Technology vol 67 pp 138ndash147 2014
[19] C Li R Huang Z Ding et al ldquoA level set method for imagesegmentation in the presence of intensity inhomogeneities withapplication toMRIrdquo IEEE Transactions on Image Processing vol20 no 7 pp 2007ndash2016 2011
[20] C Li J C Gore and C Davatzikos ldquoMultiplicative intrinsiccomponent optimization (MICO) for MRI bias field estimationand tissue segmentationrdquoMagnetic Resonance Imaging vol 32no 7 pp 913ndash923 2014
[21] A Vahadane T Peng A Sethi et al ldquoStructure-preservingcolor normalization and sparse stain separation for histologicalimagesrdquo IEEE Transactions on Medical Imaging vol 35 no 8pp 1962ndash1971 2016
[22] E Reinhard M Ashikhmin B Gooch and P Shirley ldquoColortransfer between imagesrdquo IEEE Computer Graphics amp Applica-tions vol 21 no 5 pp 34ndash41 2001
[23] D L Donoho ldquoDe-noising by soft-thresholdingrdquo IEEE Trans-actions on Information Theory vol 41 no 3 pp 613ndash627 1995
[24] E Zhang H Yang and M Xu ldquoA novel tone mappingmethod for high dynamic range image by incorporating edge-preserving filter into method based on retinexrdquo Applied Mathe-matics amp Information Sciences vol 9 no 1 pp 411ndash417 2015
[25] M Macenko M Niethammer J S Marron et al ldquoA methodfor normalizing histology slides for quantitative analysisrdquo inProceedings of the IEEE International Conference on Symposiumon Biomedical Imaging From Nano To Macro pp 1107ndash1110IEEE Press 2009
[26] T Mitsunaga and S K Nayar ldquoRadiometric self calibrationrdquoin Proceedings of the IEEE Computer Society Conference onComputer Vision and Pattern Recognition vol 1 p 1374 FortCollins Colo USA June 1999
[27] F Durand and J Dorsey ldquoFast bilateral filtering for the displayof high-dynamic-range imagesrdquoACMTransactions onGraphicsvol 21 no 3 pp 257ndash266 2002
[28] D J Jobson Z-U Rahman andG AWoodell ldquoProperties andperformance of a centersurround retinexrdquo IEEE Transactionson Image Processing vol 6 no 3 pp 451ndash462 1997
[29] D J Jobson Z-U Rahman and G A Woodell ldquoA multiscaleretinex for bridging the gap between color images and thehuman observation of scenesrdquo IEEE Transactions on ImageProcessing vol 6 no 7 pp 965ndash976 1997
BioMed Research International 11
[30] YWangQ Chen andB Zhang ldquoImage enhancement based onequal area dualistic sub-image histogram equalizationmethodrdquoIEEE Transactions on Consumer Electronics vol 45 no 1 pp68ndash75 1999
[31] S-D Chen and A R Ramli ldquoMinimum mean brightnesserror bi-histogram equalization in contrast enhancementrdquo IEEETransactions on Consumer Electronics vol 49 no 4 pp 1310ndash1319 2003
[32] S-D Chen and A R Ramli ldquoContrast enhancement usingrecursive mean-separate histogram equalization for scalablebrightness preservationrdquo IEEE Transactions on Consumer Elec-tronics vol 49 no 4 pp 1301ndash1309 2003
[33] K S Sim C P Tso and Y Y Tan ldquoRecursive sub-imagehistogram equalization applied to gray scale imagesrdquo PatternRecognition Letters vol 28 no 10 pp 1209ndash1221 2007
[34] V L Jaya and R Gopikakumari ldquoIEM a new image enhance-ment metric for contrast and sharpness measurementsrdquo Inter-national Journal of Computer Applications vol 79 no 9 pp 1ndash92013
[35] S S Agaian K Panetta and A M Grigoryan ldquoTransform-based image enhancement algorithms with performance mea-surerdquo IEEE Transactions on Image Processing vol 10 no 3 pp367ndash382 2001
[36] M Braiki A Benzinou K Nasreddine S Labidi and NHymery ldquoSegmentation of dendritic cells from microscopicimages using mathematical morphologyrdquo in Proceedings ofthe 2nd International Conference on Advanced Technologies forSignal and Image Processing (ATSIP rsquo16) pp 282ndash287 MonastirTunisia March 2016
[37] L R Dice ldquoMeasures of the amount of ecologic associationbetween speciesrdquo Ecology vol 26 no 3 pp 297ndash302 1945
[38] P Jaccard ldquoThe distribution of the flora in the alpine zonerdquoNewPhytologist vol 11 no 2 pp 37ndash50 1912
(a) Original image (b) Reference image for stain normalization
(c) Normalized image (d) Corrected image by improved bias field model
(e) E channel image (f) H channel image
(g) HDR image after tone mapping (h) Detail enhanced HDR image
Figure 4 The result of our proposed HDR pathological image enhancement method
image by the above comparison algorithms respectively andFigure 5(i) is our result
We can see from Figure 5 that all the above methodscan enhance the pathological images to some extent But the
differences between the results are a little big To evaluatecorrectly the enhancement results of different algorithmsthis paper analyze the corresponding images from both thesubjective and objective aspects
BioMed Research International 7
(a) Original image (b) Frankle-McCann Retinex (c) SSR
(d) MSR (e) DSIHE (f) MMBEBHE
(g) RMSHE (h) RSIHE (i) Proposed
Figure 5 The comparison of our method with other different image enhancement methods
Table 2 The number of best enhanced image selected by pathologists for different algorithms
41 Subjective Evaluation Pathological image is the goldstandard in disease diagnosis In order to verify the prac-ticability of our algorithm we extract 50 groups of imagesrandomly and invite three pathologists to choose one bestimage from every group The selected image should be theoptimal in visual quality and the most helpful to clinicalapplication from the perspective of a pathologist The finalresult is shown in Table 2
As Table 2 shows the results of HE-based algorithmsare basically not recognized by pathologists because there isserious distortion problem in the enhanced image of thosealgorithms Most of the detail information in those images islost and the cell regions cannot be distinguished Although itis possible to obtain the optimal enhancement effect for everyone of the other algorithms our proposed method is betterstatistically having the maximum amount of the selected best
8 BioMed Research International
Table 3 Data comparison of our method with other different image enhancement methods
Algorithms EvaluationsPSNR (dB) SD Mean EME Entropy (bit) Run time (second)
image So from a pathologistrsquos viewpoint our pathologicalimage enhancement method can improve the image qualitya lot
42 Objective Evaluation On the other hand we also analyzethe experiment results quantitatively in five well-known met-rics namely peak signal noise ratio (PSNR) [34] standarddeviation (SD) mean measure of enhancement (EME) [35]and information entropy [34]
PSNR is widely used to evaluate the quality of imageThehigher PSNR value denotes that the image could suppressnoise better The calculation equation of PSNR is as follows
and where 119883 and 119884 are input image and output imagerespectively119873 is the total number of image pixels and119871 is thedynamic range of pixel values The unit of PSNR is decibels(dB)
Themean value is used to evaluate the average luminanceof image The higher mean value represents that the image isbrighter SD is also a popular metric in image enhancementand is used to estimate the contrast of image The imagecontrast is grater if the SD value is higher For mean and SDwe compute the average value of three channels of the colorpathological image in our experiments
EME is one well-known blind-reference image qualityassessment (IQA)metric It gives a quality score to each imagebased on the image contrast The larger EME value representsthe more detail information and more obvious variation inlocal region The definition of EME is as follows
EME11989611198962 =1
1198961 11989621198962
sum119897=1
1198961
sum119896=1
20 log 119868120596max119896119897
119868120596min119896119897 (13)
where the test image is divided into 1198961 times 1198962 small blocksand 119868120596max119896119897 and 119868120596min119896119897 represent the maximum andminimumvalues of pixel respectively in block 120596119896119897
Information entropy is an important metric to measurethe content of image And the higher value indicates an imagewith richer details The equation of entropy is as follows
119867 = minus119872
sum119896=1
119901119896log2119901119896 (14)
where119872 is the gray levels of image and 119901119896 is the probabilityof gray level 119896 in the whole image
All the objective evaluation result data is shown in Table 3and all the data is an average value of our 140 pathologicalimages We can see from Table 3 that all these algorithmscould enhance image to some extent The SD values of HE-based methods are much higher than other methods Buttheir contribution is not outstanding when taking the seriousdistortion problem of those enhanced images into accountIn addition the PSNR value of our algorithm is not as goodas the Retinex-based algorithms but we obtain the bestresults in all the other metrics of mean EME and entropyAnd considering that we just use the original image andthe enhanced image to replace the noise-free image and testimage in the definition of PSNR the contribution of PSNR islimited So in conclusion our method is superior to others
Moreover in order to indicate the complexity of ouralgorithm we compare the run times of all the algorithmsrunning in the same platform and using the same imageas described above The final average result is shown in thelast column of Table 3 It is obvious that our algorithm ismuch faster than Retinex-based algorithms Although therun time of HE-based algorithm is the least the time cost ofour algorithm is worthy in view of the quality of enhancedimages
43 Cell Segmentation We also use a simple automaticcell segmentation method based on morphology to ver-ify that our image enhancement algorithm can improvesubsequent image segmentation and other image analysisprocesses In this cell segmentation method firstly transformthe original RGB image into gray-scale image and thenconduct image opening operation image reconstructionimage binarization image erosion and dilation operationsand image denoising [36] We select 20 groups of enhanced
BioMed Research International 9
(a) Original image (b) Ground truth (c) Frankle-McCann (d) SSR (e) MSR
images in which the cell regions are relatively obvious to dotest Figure 6 demonstrates one group segmentation resultwhere Figure 6(a) is the original RGB pathological imageFigure 6(b) is the ground truth segmented manually underthe guidance of pathologists and Figure 6(c)ndashFigure 6(j)are obtained from segmenting the corresponding enhancedimages
It is clear that the cell segmentation result of ouralgorithm is more close to the ground truth In order tocompare the segmentation results quantitatively we adopttwo standard segmentation metrics to evaluate namely dicecoefficient [37] and Jaccard index [38] The definitions ofthem are as follows
DC (SRTR) = 2 timesNum (pixelSR cap pixelTR)Num (pixelSR) +Num (pixelTR)
JI (SRTR) = Num (pixelSR cap pixelTR)Num (pixelSR cup pixelTR)
(15)
where SR is the segmentation region TR is the true regionof the target and Num(pixel) is the number of relevantpixel points The average results of performance metricsare displayed in Table 4 We can see that the result ofour proposed method is the best Our HDR pathologicalimage enhancement method is superior to the comparisonalgorithms as we can improve the image segmentation andpathological analysis better
In sum we proposed that HDR pathological imageenhancement method obtains a better result according to
both pathologistsrsquo subjective evaluation and quantitativeanalysis in data and the cell segmentationmethod also provesthe better performance of our method as the quality anddetail of original image are both improved
5 Conclusions
This paper proposes new HDR pathological image enhance-ment methods based on GIF and improved bias fieldcorrection model First stain normalization and waveletdenoising operations are used in image preprocessing Andthe improved bias field model is introduced to correct theintensity inhomogeneity and detail discontinuity of imageThen the HDR pathological image is generated using LDRimage and H and E channel images Next the 119884 componentof HDR image is separated into base layer and detail layerby GIF and the two layers are enhanced separately Finallythe fine enhanced image is acquired after combining the 119884component and the color components To verify the effective-ness of the proposed method we perform the enhancementexperiments using 140 pathological images The experimentresults and comparisons with related work demonstrate thatour proposed method improves the image quality in terms ofhuman vision PSNR SD mean EME information entropyand cell segmentation
Competing Interests
The authors declare that there were no competing interestsregarding the publication of this article
10 BioMed Research International
Acknowledgments
This work was supported by the National Natural ScienceFoundation of China (no 61472073)
References
[1] SM Pizer E P Amburn J D Austin et al ldquoAdaptive histogramequalization and its variationsrdquo Computer Vision Graphics ampImage Processing vol 39 no 3 pp 355ndash368 1987
[2] V Vani and K V M Prashanth ldquoColor image enhancementtechniques in Wireless Capsule Endoscopyrdquo in Proceedings ofthe IEEE International Conference on Trends in AutomationCommunications and Computing Technology (I-TACT rsquo15) vol1 pp 1ndash6 Bangalore India December 2015
[3] H Cao L Tian J Liu H Wang and S Feng ldquoColor imageenhancement using power-constraint histogram equalizationfor AMOLEDrdquo in Proceedings of the IEEE 11th InternationalConference on ASIC (ASICON rsquo15) pp 1ndash4 IEEE ChengduChina November 2015
[4] N M Kwok G Fang and H Y Shi ldquoColor enhancementfor images from digital camera using a transformation-freeapproachrdquo in Proceedings of the 9th International Conferenceon Sensing Technology (ICST rsquo15) pp 168ndash172 IEEE AucklandNew Zealand December 2015
[5] S D Nikam and R U Yawale ldquoColor image enhancementusing daubechies wavelet transform and HIS color modelrdquoin Proceedings of the International Conference on IndustrialInstrumentation and Control (ICIC rsquo15) pp 1323ndash1327 IEEEPune India May 2015
[6] L G Villanueva G M Callico F Tobajas et al ldquoMedicaldiagnosis improvement through image quality enhancementbased on super-resolutionrdquo in Proceedings of the 13th EuromicroConference onDigital SystemDesign ArchitecturesMethods andTools (DSD rsquo10) pp 259ndash262 IEEE Lille France September2010
[7] W Sun F Li and Q Zhang ldquoThe applications of improvedretinex algorithm for X-ray medical image enhancementrdquo inProceedings of the International Conference on Computer Scienceand Service System (CSSS rsquo12) pp 1655ndash1658 IEEE NanjingChina August 2012
[8] G Zhang D Sun P Yan H Zhao and Z Li ldquoA LDCT imagecontrast enhancement algorithm based on single-scale retinextheoryrdquo in Proceedings of the International Conference on Com-putational Intelligence for Modelling Control amp Automation pp1282ndash1287 IEEE Computer Society Vienna Austria December2008
[9] S Setty N K Srinath and M C Hanumantharaju ldquoDevel-opment of multiscale retinex algorithm for medical imageenhancement based on multi-rate samplingrdquo in Proceedingsof the International Conference on Signal Processing ImageProcessing amp Pattern Recognition pp 145ndash150 2013
[10] J Mccann ldquoLessons learned from mondrians applied to realimages and color gamutsrdquo in Proceedings of the Color andImaging Conference vol 8 pp 1ndash8 1999
[11] B V Funt F Ciurea and J J McCann ldquoRetinex in Matlabrdquo inProceedings of the Color and Imaging Conference pp 112ndash121Scottsdale Ariz USA November 2000
[12] K Kim J Bae and J Kim ldquoNatural hdr image tone mappingbased on retinexrdquo IEEE Transactions on Consumer Electronicsvol 57 no 4 pp 1807ndash1814 2011
[13] J Kuang G M Johnson and M D Fairchild ldquoiCAM06 arefined image appearance model for HDR image renderingrdquoJournal of Visual Communication amp Image Representation vol18 no 5 pp 406ndash414 2007
[14] M-L Song H-Q Wang C Chen X-Q Ye and W-K GuldquoTone mapping for high dynamic range image using a proba-bilistic modelrdquo Journal of Software vol 20 no 3 pp 734ndash7432010
[15] F Branchitta M Diani G Corsini and M Romagnoli ldquoNewtechnique for the visualization of high dynamic range infraredimagesrdquo Optical Engineering vol 48 no 9 Article ID 0964012009
[16] C Zuo ldquoDisplay and detail enhancement for high-dynamic-range infrared imagesrdquo Optical Engineering vol 50 no 12Article ID 127401 pp 895ndash900 2011
[17] K He J Sun and X Tang ldquoGuided image filteringrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol35 no 6 pp 1397ndash1409 2013
[18] N Liu and D Zhao ldquoDetail enhancement for high-dynamic-range infrared images based on guided image filterrdquo InfraredPhysics amp Technology vol 67 pp 138ndash147 2014
[19] C Li R Huang Z Ding et al ldquoA level set method for imagesegmentation in the presence of intensity inhomogeneities withapplication toMRIrdquo IEEE Transactions on Image Processing vol20 no 7 pp 2007ndash2016 2011
[20] C Li J C Gore and C Davatzikos ldquoMultiplicative intrinsiccomponent optimization (MICO) for MRI bias field estimationand tissue segmentationrdquoMagnetic Resonance Imaging vol 32no 7 pp 913ndash923 2014
[21] A Vahadane T Peng A Sethi et al ldquoStructure-preservingcolor normalization and sparse stain separation for histologicalimagesrdquo IEEE Transactions on Medical Imaging vol 35 no 8pp 1962ndash1971 2016
[22] E Reinhard M Ashikhmin B Gooch and P Shirley ldquoColortransfer between imagesrdquo IEEE Computer Graphics amp Applica-tions vol 21 no 5 pp 34ndash41 2001
[23] D L Donoho ldquoDe-noising by soft-thresholdingrdquo IEEE Trans-actions on Information Theory vol 41 no 3 pp 613ndash627 1995
[24] E Zhang H Yang and M Xu ldquoA novel tone mappingmethod for high dynamic range image by incorporating edge-preserving filter into method based on retinexrdquo Applied Mathe-matics amp Information Sciences vol 9 no 1 pp 411ndash417 2015
[25] M Macenko M Niethammer J S Marron et al ldquoA methodfor normalizing histology slides for quantitative analysisrdquo inProceedings of the IEEE International Conference on Symposiumon Biomedical Imaging From Nano To Macro pp 1107ndash1110IEEE Press 2009
[26] T Mitsunaga and S K Nayar ldquoRadiometric self calibrationrdquoin Proceedings of the IEEE Computer Society Conference onComputer Vision and Pattern Recognition vol 1 p 1374 FortCollins Colo USA June 1999
[27] F Durand and J Dorsey ldquoFast bilateral filtering for the displayof high-dynamic-range imagesrdquoACMTransactions onGraphicsvol 21 no 3 pp 257ndash266 2002
[28] D J Jobson Z-U Rahman andG AWoodell ldquoProperties andperformance of a centersurround retinexrdquo IEEE Transactionson Image Processing vol 6 no 3 pp 451ndash462 1997
[29] D J Jobson Z-U Rahman and G A Woodell ldquoA multiscaleretinex for bridging the gap between color images and thehuman observation of scenesrdquo IEEE Transactions on ImageProcessing vol 6 no 7 pp 965ndash976 1997
BioMed Research International 11
[30] YWangQ Chen andB Zhang ldquoImage enhancement based onequal area dualistic sub-image histogram equalizationmethodrdquoIEEE Transactions on Consumer Electronics vol 45 no 1 pp68ndash75 1999
[31] S-D Chen and A R Ramli ldquoMinimum mean brightnesserror bi-histogram equalization in contrast enhancementrdquo IEEETransactions on Consumer Electronics vol 49 no 4 pp 1310ndash1319 2003
[32] S-D Chen and A R Ramli ldquoContrast enhancement usingrecursive mean-separate histogram equalization for scalablebrightness preservationrdquo IEEE Transactions on Consumer Elec-tronics vol 49 no 4 pp 1301ndash1309 2003
[33] K S Sim C P Tso and Y Y Tan ldquoRecursive sub-imagehistogram equalization applied to gray scale imagesrdquo PatternRecognition Letters vol 28 no 10 pp 1209ndash1221 2007
[34] V L Jaya and R Gopikakumari ldquoIEM a new image enhance-ment metric for contrast and sharpness measurementsrdquo Inter-national Journal of Computer Applications vol 79 no 9 pp 1ndash92013
[35] S S Agaian K Panetta and A M Grigoryan ldquoTransform-based image enhancement algorithms with performance mea-surerdquo IEEE Transactions on Image Processing vol 10 no 3 pp367ndash382 2001
[36] M Braiki A Benzinou K Nasreddine S Labidi and NHymery ldquoSegmentation of dendritic cells from microscopicimages using mathematical morphologyrdquo in Proceedings ofthe 2nd International Conference on Advanced Technologies forSignal and Image Processing (ATSIP rsquo16) pp 282ndash287 MonastirTunisia March 2016
[37] L R Dice ldquoMeasures of the amount of ecologic associationbetween speciesrdquo Ecology vol 26 no 3 pp 297ndash302 1945
[38] P Jaccard ldquoThe distribution of the flora in the alpine zonerdquoNewPhytologist vol 11 no 2 pp 37ndash50 1912
41 Subjective Evaluation Pathological image is the goldstandard in disease diagnosis In order to verify the prac-ticability of our algorithm we extract 50 groups of imagesrandomly and invite three pathologists to choose one bestimage from every group The selected image should be theoptimal in visual quality and the most helpful to clinicalapplication from the perspective of a pathologist The finalresult is shown in Table 2
As Table 2 shows the results of HE-based algorithmsare basically not recognized by pathologists because there isserious distortion problem in the enhanced image of thosealgorithms Most of the detail information in those images islost and the cell regions cannot be distinguished Although itis possible to obtain the optimal enhancement effect for everyone of the other algorithms our proposed method is betterstatistically having the maximum amount of the selected best
8 BioMed Research International
Table 3 Data comparison of our method with other different image enhancement methods
Algorithms EvaluationsPSNR (dB) SD Mean EME Entropy (bit) Run time (second)
image So from a pathologistrsquos viewpoint our pathologicalimage enhancement method can improve the image qualitya lot
42 Objective Evaluation On the other hand we also analyzethe experiment results quantitatively in five well-known met-rics namely peak signal noise ratio (PSNR) [34] standarddeviation (SD) mean measure of enhancement (EME) [35]and information entropy [34]
PSNR is widely used to evaluate the quality of imageThehigher PSNR value denotes that the image could suppressnoise better The calculation equation of PSNR is as follows
and where 119883 and 119884 are input image and output imagerespectively119873 is the total number of image pixels and119871 is thedynamic range of pixel values The unit of PSNR is decibels(dB)
Themean value is used to evaluate the average luminanceof image The higher mean value represents that the image isbrighter SD is also a popular metric in image enhancementand is used to estimate the contrast of image The imagecontrast is grater if the SD value is higher For mean and SDwe compute the average value of three channels of the colorpathological image in our experiments
EME is one well-known blind-reference image qualityassessment (IQA)metric It gives a quality score to each imagebased on the image contrast The larger EME value representsthe more detail information and more obvious variation inlocal region The definition of EME is as follows
EME11989611198962 =1
1198961 11989621198962
sum119897=1
1198961
sum119896=1
20 log 119868120596max119896119897
119868120596min119896119897 (13)
where the test image is divided into 1198961 times 1198962 small blocksand 119868120596max119896119897 and 119868120596min119896119897 represent the maximum andminimumvalues of pixel respectively in block 120596119896119897
Information entropy is an important metric to measurethe content of image And the higher value indicates an imagewith richer details The equation of entropy is as follows
119867 = minus119872
sum119896=1
119901119896log2119901119896 (14)
where119872 is the gray levels of image and 119901119896 is the probabilityof gray level 119896 in the whole image
All the objective evaluation result data is shown in Table 3and all the data is an average value of our 140 pathologicalimages We can see from Table 3 that all these algorithmscould enhance image to some extent The SD values of HE-based methods are much higher than other methods Buttheir contribution is not outstanding when taking the seriousdistortion problem of those enhanced images into accountIn addition the PSNR value of our algorithm is not as goodas the Retinex-based algorithms but we obtain the bestresults in all the other metrics of mean EME and entropyAnd considering that we just use the original image andthe enhanced image to replace the noise-free image and testimage in the definition of PSNR the contribution of PSNR islimited So in conclusion our method is superior to others
Moreover in order to indicate the complexity of ouralgorithm we compare the run times of all the algorithmsrunning in the same platform and using the same imageas described above The final average result is shown in thelast column of Table 3 It is obvious that our algorithm ismuch faster than Retinex-based algorithms Although therun time of HE-based algorithm is the least the time cost ofour algorithm is worthy in view of the quality of enhancedimages
43 Cell Segmentation We also use a simple automaticcell segmentation method based on morphology to ver-ify that our image enhancement algorithm can improvesubsequent image segmentation and other image analysisprocesses In this cell segmentation method firstly transformthe original RGB image into gray-scale image and thenconduct image opening operation image reconstructionimage binarization image erosion and dilation operationsand image denoising [36] We select 20 groups of enhanced
BioMed Research International 9
(a) Original image (b) Ground truth (c) Frankle-McCann (d) SSR (e) MSR
images in which the cell regions are relatively obvious to dotest Figure 6 demonstrates one group segmentation resultwhere Figure 6(a) is the original RGB pathological imageFigure 6(b) is the ground truth segmented manually underthe guidance of pathologists and Figure 6(c)ndashFigure 6(j)are obtained from segmenting the corresponding enhancedimages
It is clear that the cell segmentation result of ouralgorithm is more close to the ground truth In order tocompare the segmentation results quantitatively we adopttwo standard segmentation metrics to evaluate namely dicecoefficient [37] and Jaccard index [38] The definitions ofthem are as follows
DC (SRTR) = 2 timesNum (pixelSR cap pixelTR)Num (pixelSR) +Num (pixelTR)
JI (SRTR) = Num (pixelSR cap pixelTR)Num (pixelSR cup pixelTR)
(15)
where SR is the segmentation region TR is the true regionof the target and Num(pixel) is the number of relevantpixel points The average results of performance metricsare displayed in Table 4 We can see that the result ofour proposed method is the best Our HDR pathologicalimage enhancement method is superior to the comparisonalgorithms as we can improve the image segmentation andpathological analysis better
In sum we proposed that HDR pathological imageenhancement method obtains a better result according to
both pathologistsrsquo subjective evaluation and quantitativeanalysis in data and the cell segmentationmethod also provesthe better performance of our method as the quality anddetail of original image are both improved
5 Conclusions
This paper proposes new HDR pathological image enhance-ment methods based on GIF and improved bias fieldcorrection model First stain normalization and waveletdenoising operations are used in image preprocessing Andthe improved bias field model is introduced to correct theintensity inhomogeneity and detail discontinuity of imageThen the HDR pathological image is generated using LDRimage and H and E channel images Next the 119884 componentof HDR image is separated into base layer and detail layerby GIF and the two layers are enhanced separately Finallythe fine enhanced image is acquired after combining the 119884component and the color components To verify the effective-ness of the proposed method we perform the enhancementexperiments using 140 pathological images The experimentresults and comparisons with related work demonstrate thatour proposed method improves the image quality in terms ofhuman vision PSNR SD mean EME information entropyand cell segmentation
Competing Interests
The authors declare that there were no competing interestsregarding the publication of this article
10 BioMed Research International
Acknowledgments
This work was supported by the National Natural ScienceFoundation of China (no 61472073)
References
[1] SM Pizer E P Amburn J D Austin et al ldquoAdaptive histogramequalization and its variationsrdquo Computer Vision Graphics ampImage Processing vol 39 no 3 pp 355ndash368 1987
[2] V Vani and K V M Prashanth ldquoColor image enhancementtechniques in Wireless Capsule Endoscopyrdquo in Proceedings ofthe IEEE International Conference on Trends in AutomationCommunications and Computing Technology (I-TACT rsquo15) vol1 pp 1ndash6 Bangalore India December 2015
[3] H Cao L Tian J Liu H Wang and S Feng ldquoColor imageenhancement using power-constraint histogram equalizationfor AMOLEDrdquo in Proceedings of the IEEE 11th InternationalConference on ASIC (ASICON rsquo15) pp 1ndash4 IEEE ChengduChina November 2015
[4] N M Kwok G Fang and H Y Shi ldquoColor enhancementfor images from digital camera using a transformation-freeapproachrdquo in Proceedings of the 9th International Conferenceon Sensing Technology (ICST rsquo15) pp 168ndash172 IEEE AucklandNew Zealand December 2015
[5] S D Nikam and R U Yawale ldquoColor image enhancementusing daubechies wavelet transform and HIS color modelrdquoin Proceedings of the International Conference on IndustrialInstrumentation and Control (ICIC rsquo15) pp 1323ndash1327 IEEEPune India May 2015
[6] L G Villanueva G M Callico F Tobajas et al ldquoMedicaldiagnosis improvement through image quality enhancementbased on super-resolutionrdquo in Proceedings of the 13th EuromicroConference onDigital SystemDesign ArchitecturesMethods andTools (DSD rsquo10) pp 259ndash262 IEEE Lille France September2010
[7] W Sun F Li and Q Zhang ldquoThe applications of improvedretinex algorithm for X-ray medical image enhancementrdquo inProceedings of the International Conference on Computer Scienceand Service System (CSSS rsquo12) pp 1655ndash1658 IEEE NanjingChina August 2012
[8] G Zhang D Sun P Yan H Zhao and Z Li ldquoA LDCT imagecontrast enhancement algorithm based on single-scale retinextheoryrdquo in Proceedings of the International Conference on Com-putational Intelligence for Modelling Control amp Automation pp1282ndash1287 IEEE Computer Society Vienna Austria December2008
[9] S Setty N K Srinath and M C Hanumantharaju ldquoDevel-opment of multiscale retinex algorithm for medical imageenhancement based on multi-rate samplingrdquo in Proceedingsof the International Conference on Signal Processing ImageProcessing amp Pattern Recognition pp 145ndash150 2013
[10] J Mccann ldquoLessons learned from mondrians applied to realimages and color gamutsrdquo in Proceedings of the Color andImaging Conference vol 8 pp 1ndash8 1999
[11] B V Funt F Ciurea and J J McCann ldquoRetinex in Matlabrdquo inProceedings of the Color and Imaging Conference pp 112ndash121Scottsdale Ariz USA November 2000
[12] K Kim J Bae and J Kim ldquoNatural hdr image tone mappingbased on retinexrdquo IEEE Transactions on Consumer Electronicsvol 57 no 4 pp 1807ndash1814 2011
[13] J Kuang G M Johnson and M D Fairchild ldquoiCAM06 arefined image appearance model for HDR image renderingrdquoJournal of Visual Communication amp Image Representation vol18 no 5 pp 406ndash414 2007
[14] M-L Song H-Q Wang C Chen X-Q Ye and W-K GuldquoTone mapping for high dynamic range image using a proba-bilistic modelrdquo Journal of Software vol 20 no 3 pp 734ndash7432010
[15] F Branchitta M Diani G Corsini and M Romagnoli ldquoNewtechnique for the visualization of high dynamic range infraredimagesrdquo Optical Engineering vol 48 no 9 Article ID 0964012009
[16] C Zuo ldquoDisplay and detail enhancement for high-dynamic-range infrared imagesrdquo Optical Engineering vol 50 no 12Article ID 127401 pp 895ndash900 2011
[17] K He J Sun and X Tang ldquoGuided image filteringrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol35 no 6 pp 1397ndash1409 2013
[18] N Liu and D Zhao ldquoDetail enhancement for high-dynamic-range infrared images based on guided image filterrdquo InfraredPhysics amp Technology vol 67 pp 138ndash147 2014
[19] C Li R Huang Z Ding et al ldquoA level set method for imagesegmentation in the presence of intensity inhomogeneities withapplication toMRIrdquo IEEE Transactions on Image Processing vol20 no 7 pp 2007ndash2016 2011
[20] C Li J C Gore and C Davatzikos ldquoMultiplicative intrinsiccomponent optimization (MICO) for MRI bias field estimationand tissue segmentationrdquoMagnetic Resonance Imaging vol 32no 7 pp 913ndash923 2014
[21] A Vahadane T Peng A Sethi et al ldquoStructure-preservingcolor normalization and sparse stain separation for histologicalimagesrdquo IEEE Transactions on Medical Imaging vol 35 no 8pp 1962ndash1971 2016
[22] E Reinhard M Ashikhmin B Gooch and P Shirley ldquoColortransfer between imagesrdquo IEEE Computer Graphics amp Applica-tions vol 21 no 5 pp 34ndash41 2001
[23] D L Donoho ldquoDe-noising by soft-thresholdingrdquo IEEE Trans-actions on Information Theory vol 41 no 3 pp 613ndash627 1995
[24] E Zhang H Yang and M Xu ldquoA novel tone mappingmethod for high dynamic range image by incorporating edge-preserving filter into method based on retinexrdquo Applied Mathe-matics amp Information Sciences vol 9 no 1 pp 411ndash417 2015
[25] M Macenko M Niethammer J S Marron et al ldquoA methodfor normalizing histology slides for quantitative analysisrdquo inProceedings of the IEEE International Conference on Symposiumon Biomedical Imaging From Nano To Macro pp 1107ndash1110IEEE Press 2009
[26] T Mitsunaga and S K Nayar ldquoRadiometric self calibrationrdquoin Proceedings of the IEEE Computer Society Conference onComputer Vision and Pattern Recognition vol 1 p 1374 FortCollins Colo USA June 1999
[27] F Durand and J Dorsey ldquoFast bilateral filtering for the displayof high-dynamic-range imagesrdquoACMTransactions onGraphicsvol 21 no 3 pp 257ndash266 2002
[28] D J Jobson Z-U Rahman andG AWoodell ldquoProperties andperformance of a centersurround retinexrdquo IEEE Transactionson Image Processing vol 6 no 3 pp 451ndash462 1997
[29] D J Jobson Z-U Rahman and G A Woodell ldquoA multiscaleretinex for bridging the gap between color images and thehuman observation of scenesrdquo IEEE Transactions on ImageProcessing vol 6 no 7 pp 965ndash976 1997
BioMed Research International 11
[30] YWangQ Chen andB Zhang ldquoImage enhancement based onequal area dualistic sub-image histogram equalizationmethodrdquoIEEE Transactions on Consumer Electronics vol 45 no 1 pp68ndash75 1999
[31] S-D Chen and A R Ramli ldquoMinimum mean brightnesserror bi-histogram equalization in contrast enhancementrdquo IEEETransactions on Consumer Electronics vol 49 no 4 pp 1310ndash1319 2003
[32] S-D Chen and A R Ramli ldquoContrast enhancement usingrecursive mean-separate histogram equalization for scalablebrightness preservationrdquo IEEE Transactions on Consumer Elec-tronics vol 49 no 4 pp 1301ndash1309 2003
[33] K S Sim C P Tso and Y Y Tan ldquoRecursive sub-imagehistogram equalization applied to gray scale imagesrdquo PatternRecognition Letters vol 28 no 10 pp 1209ndash1221 2007
[34] V L Jaya and R Gopikakumari ldquoIEM a new image enhance-ment metric for contrast and sharpness measurementsrdquo Inter-national Journal of Computer Applications vol 79 no 9 pp 1ndash92013
[35] S S Agaian K Panetta and A M Grigoryan ldquoTransform-based image enhancement algorithms with performance mea-surerdquo IEEE Transactions on Image Processing vol 10 no 3 pp367ndash382 2001
[36] M Braiki A Benzinou K Nasreddine S Labidi and NHymery ldquoSegmentation of dendritic cells from microscopicimages using mathematical morphologyrdquo in Proceedings ofthe 2nd International Conference on Advanced Technologies forSignal and Image Processing (ATSIP rsquo16) pp 282ndash287 MonastirTunisia March 2016
[37] L R Dice ldquoMeasures of the amount of ecologic associationbetween speciesrdquo Ecology vol 26 no 3 pp 297ndash302 1945
[38] P Jaccard ldquoThe distribution of the flora in the alpine zonerdquoNewPhytologist vol 11 no 2 pp 37ndash50 1912
image So from a pathologistrsquos viewpoint our pathologicalimage enhancement method can improve the image qualitya lot
42 Objective Evaluation On the other hand we also analyzethe experiment results quantitatively in five well-known met-rics namely peak signal noise ratio (PSNR) [34] standarddeviation (SD) mean measure of enhancement (EME) [35]and information entropy [34]
PSNR is widely used to evaluate the quality of imageThehigher PSNR value denotes that the image could suppressnoise better The calculation equation of PSNR is as follows
and where 119883 and 119884 are input image and output imagerespectively119873 is the total number of image pixels and119871 is thedynamic range of pixel values The unit of PSNR is decibels(dB)
Themean value is used to evaluate the average luminanceof image The higher mean value represents that the image isbrighter SD is also a popular metric in image enhancementand is used to estimate the contrast of image The imagecontrast is grater if the SD value is higher For mean and SDwe compute the average value of three channels of the colorpathological image in our experiments
EME is one well-known blind-reference image qualityassessment (IQA)metric It gives a quality score to each imagebased on the image contrast The larger EME value representsthe more detail information and more obvious variation inlocal region The definition of EME is as follows
EME11989611198962 =1
1198961 11989621198962
sum119897=1
1198961
sum119896=1
20 log 119868120596max119896119897
119868120596min119896119897 (13)
where the test image is divided into 1198961 times 1198962 small blocksand 119868120596max119896119897 and 119868120596min119896119897 represent the maximum andminimumvalues of pixel respectively in block 120596119896119897
Information entropy is an important metric to measurethe content of image And the higher value indicates an imagewith richer details The equation of entropy is as follows
119867 = minus119872
sum119896=1
119901119896log2119901119896 (14)
where119872 is the gray levels of image and 119901119896 is the probabilityof gray level 119896 in the whole image
All the objective evaluation result data is shown in Table 3and all the data is an average value of our 140 pathologicalimages We can see from Table 3 that all these algorithmscould enhance image to some extent The SD values of HE-based methods are much higher than other methods Buttheir contribution is not outstanding when taking the seriousdistortion problem of those enhanced images into accountIn addition the PSNR value of our algorithm is not as goodas the Retinex-based algorithms but we obtain the bestresults in all the other metrics of mean EME and entropyAnd considering that we just use the original image andthe enhanced image to replace the noise-free image and testimage in the definition of PSNR the contribution of PSNR islimited So in conclusion our method is superior to others
Moreover in order to indicate the complexity of ouralgorithm we compare the run times of all the algorithmsrunning in the same platform and using the same imageas described above The final average result is shown in thelast column of Table 3 It is obvious that our algorithm ismuch faster than Retinex-based algorithms Although therun time of HE-based algorithm is the least the time cost ofour algorithm is worthy in view of the quality of enhancedimages
43 Cell Segmentation We also use a simple automaticcell segmentation method based on morphology to ver-ify that our image enhancement algorithm can improvesubsequent image segmentation and other image analysisprocesses In this cell segmentation method firstly transformthe original RGB image into gray-scale image and thenconduct image opening operation image reconstructionimage binarization image erosion and dilation operationsand image denoising [36] We select 20 groups of enhanced
BioMed Research International 9
(a) Original image (b) Ground truth (c) Frankle-McCann (d) SSR (e) MSR
images in which the cell regions are relatively obvious to dotest Figure 6 demonstrates one group segmentation resultwhere Figure 6(a) is the original RGB pathological imageFigure 6(b) is the ground truth segmented manually underthe guidance of pathologists and Figure 6(c)ndashFigure 6(j)are obtained from segmenting the corresponding enhancedimages
It is clear that the cell segmentation result of ouralgorithm is more close to the ground truth In order tocompare the segmentation results quantitatively we adopttwo standard segmentation metrics to evaluate namely dicecoefficient [37] and Jaccard index [38] The definitions ofthem are as follows
DC (SRTR) = 2 timesNum (pixelSR cap pixelTR)Num (pixelSR) +Num (pixelTR)
JI (SRTR) = Num (pixelSR cap pixelTR)Num (pixelSR cup pixelTR)
(15)
where SR is the segmentation region TR is the true regionof the target and Num(pixel) is the number of relevantpixel points The average results of performance metricsare displayed in Table 4 We can see that the result ofour proposed method is the best Our HDR pathologicalimage enhancement method is superior to the comparisonalgorithms as we can improve the image segmentation andpathological analysis better
In sum we proposed that HDR pathological imageenhancement method obtains a better result according to
both pathologistsrsquo subjective evaluation and quantitativeanalysis in data and the cell segmentationmethod also provesthe better performance of our method as the quality anddetail of original image are both improved
5 Conclusions
This paper proposes new HDR pathological image enhance-ment methods based on GIF and improved bias fieldcorrection model First stain normalization and waveletdenoising operations are used in image preprocessing Andthe improved bias field model is introduced to correct theintensity inhomogeneity and detail discontinuity of imageThen the HDR pathological image is generated using LDRimage and H and E channel images Next the 119884 componentof HDR image is separated into base layer and detail layerby GIF and the two layers are enhanced separately Finallythe fine enhanced image is acquired after combining the 119884component and the color components To verify the effective-ness of the proposed method we perform the enhancementexperiments using 140 pathological images The experimentresults and comparisons with related work demonstrate thatour proposed method improves the image quality in terms ofhuman vision PSNR SD mean EME information entropyand cell segmentation
Competing Interests
The authors declare that there were no competing interestsregarding the publication of this article
10 BioMed Research International
Acknowledgments
This work was supported by the National Natural ScienceFoundation of China (no 61472073)
References
[1] SM Pizer E P Amburn J D Austin et al ldquoAdaptive histogramequalization and its variationsrdquo Computer Vision Graphics ampImage Processing vol 39 no 3 pp 355ndash368 1987
[2] V Vani and K V M Prashanth ldquoColor image enhancementtechniques in Wireless Capsule Endoscopyrdquo in Proceedings ofthe IEEE International Conference on Trends in AutomationCommunications and Computing Technology (I-TACT rsquo15) vol1 pp 1ndash6 Bangalore India December 2015
[3] H Cao L Tian J Liu H Wang and S Feng ldquoColor imageenhancement using power-constraint histogram equalizationfor AMOLEDrdquo in Proceedings of the IEEE 11th InternationalConference on ASIC (ASICON rsquo15) pp 1ndash4 IEEE ChengduChina November 2015
[4] N M Kwok G Fang and H Y Shi ldquoColor enhancementfor images from digital camera using a transformation-freeapproachrdquo in Proceedings of the 9th International Conferenceon Sensing Technology (ICST rsquo15) pp 168ndash172 IEEE AucklandNew Zealand December 2015
[5] S D Nikam and R U Yawale ldquoColor image enhancementusing daubechies wavelet transform and HIS color modelrdquoin Proceedings of the International Conference on IndustrialInstrumentation and Control (ICIC rsquo15) pp 1323ndash1327 IEEEPune India May 2015
[6] L G Villanueva G M Callico F Tobajas et al ldquoMedicaldiagnosis improvement through image quality enhancementbased on super-resolutionrdquo in Proceedings of the 13th EuromicroConference onDigital SystemDesign ArchitecturesMethods andTools (DSD rsquo10) pp 259ndash262 IEEE Lille France September2010
[7] W Sun F Li and Q Zhang ldquoThe applications of improvedretinex algorithm for X-ray medical image enhancementrdquo inProceedings of the International Conference on Computer Scienceand Service System (CSSS rsquo12) pp 1655ndash1658 IEEE NanjingChina August 2012
[8] G Zhang D Sun P Yan H Zhao and Z Li ldquoA LDCT imagecontrast enhancement algorithm based on single-scale retinextheoryrdquo in Proceedings of the International Conference on Com-putational Intelligence for Modelling Control amp Automation pp1282ndash1287 IEEE Computer Society Vienna Austria December2008
[9] S Setty N K Srinath and M C Hanumantharaju ldquoDevel-opment of multiscale retinex algorithm for medical imageenhancement based on multi-rate samplingrdquo in Proceedingsof the International Conference on Signal Processing ImageProcessing amp Pattern Recognition pp 145ndash150 2013
[10] J Mccann ldquoLessons learned from mondrians applied to realimages and color gamutsrdquo in Proceedings of the Color andImaging Conference vol 8 pp 1ndash8 1999
[11] B V Funt F Ciurea and J J McCann ldquoRetinex in Matlabrdquo inProceedings of the Color and Imaging Conference pp 112ndash121Scottsdale Ariz USA November 2000
[12] K Kim J Bae and J Kim ldquoNatural hdr image tone mappingbased on retinexrdquo IEEE Transactions on Consumer Electronicsvol 57 no 4 pp 1807ndash1814 2011
[13] J Kuang G M Johnson and M D Fairchild ldquoiCAM06 arefined image appearance model for HDR image renderingrdquoJournal of Visual Communication amp Image Representation vol18 no 5 pp 406ndash414 2007
[14] M-L Song H-Q Wang C Chen X-Q Ye and W-K GuldquoTone mapping for high dynamic range image using a proba-bilistic modelrdquo Journal of Software vol 20 no 3 pp 734ndash7432010
[15] F Branchitta M Diani G Corsini and M Romagnoli ldquoNewtechnique for the visualization of high dynamic range infraredimagesrdquo Optical Engineering vol 48 no 9 Article ID 0964012009
[16] C Zuo ldquoDisplay and detail enhancement for high-dynamic-range infrared imagesrdquo Optical Engineering vol 50 no 12Article ID 127401 pp 895ndash900 2011
[17] K He J Sun and X Tang ldquoGuided image filteringrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol35 no 6 pp 1397ndash1409 2013
[18] N Liu and D Zhao ldquoDetail enhancement for high-dynamic-range infrared images based on guided image filterrdquo InfraredPhysics amp Technology vol 67 pp 138ndash147 2014
[19] C Li R Huang Z Ding et al ldquoA level set method for imagesegmentation in the presence of intensity inhomogeneities withapplication toMRIrdquo IEEE Transactions on Image Processing vol20 no 7 pp 2007ndash2016 2011
[20] C Li J C Gore and C Davatzikos ldquoMultiplicative intrinsiccomponent optimization (MICO) for MRI bias field estimationand tissue segmentationrdquoMagnetic Resonance Imaging vol 32no 7 pp 913ndash923 2014
[21] A Vahadane T Peng A Sethi et al ldquoStructure-preservingcolor normalization and sparse stain separation for histologicalimagesrdquo IEEE Transactions on Medical Imaging vol 35 no 8pp 1962ndash1971 2016
[22] E Reinhard M Ashikhmin B Gooch and P Shirley ldquoColortransfer between imagesrdquo IEEE Computer Graphics amp Applica-tions vol 21 no 5 pp 34ndash41 2001
[23] D L Donoho ldquoDe-noising by soft-thresholdingrdquo IEEE Trans-actions on Information Theory vol 41 no 3 pp 613ndash627 1995
[24] E Zhang H Yang and M Xu ldquoA novel tone mappingmethod for high dynamic range image by incorporating edge-preserving filter into method based on retinexrdquo Applied Mathe-matics amp Information Sciences vol 9 no 1 pp 411ndash417 2015
[25] M Macenko M Niethammer J S Marron et al ldquoA methodfor normalizing histology slides for quantitative analysisrdquo inProceedings of the IEEE International Conference on Symposiumon Biomedical Imaging From Nano To Macro pp 1107ndash1110IEEE Press 2009
[26] T Mitsunaga and S K Nayar ldquoRadiometric self calibrationrdquoin Proceedings of the IEEE Computer Society Conference onComputer Vision and Pattern Recognition vol 1 p 1374 FortCollins Colo USA June 1999
[27] F Durand and J Dorsey ldquoFast bilateral filtering for the displayof high-dynamic-range imagesrdquoACMTransactions onGraphicsvol 21 no 3 pp 257ndash266 2002
[28] D J Jobson Z-U Rahman andG AWoodell ldquoProperties andperformance of a centersurround retinexrdquo IEEE Transactionson Image Processing vol 6 no 3 pp 451ndash462 1997
[29] D J Jobson Z-U Rahman and G A Woodell ldquoA multiscaleretinex for bridging the gap between color images and thehuman observation of scenesrdquo IEEE Transactions on ImageProcessing vol 6 no 7 pp 965ndash976 1997
BioMed Research International 11
[30] YWangQ Chen andB Zhang ldquoImage enhancement based onequal area dualistic sub-image histogram equalizationmethodrdquoIEEE Transactions on Consumer Electronics vol 45 no 1 pp68ndash75 1999
[31] S-D Chen and A R Ramli ldquoMinimum mean brightnesserror bi-histogram equalization in contrast enhancementrdquo IEEETransactions on Consumer Electronics vol 49 no 4 pp 1310ndash1319 2003
[32] S-D Chen and A R Ramli ldquoContrast enhancement usingrecursive mean-separate histogram equalization for scalablebrightness preservationrdquo IEEE Transactions on Consumer Elec-tronics vol 49 no 4 pp 1301ndash1309 2003
[33] K S Sim C P Tso and Y Y Tan ldquoRecursive sub-imagehistogram equalization applied to gray scale imagesrdquo PatternRecognition Letters vol 28 no 10 pp 1209ndash1221 2007
[34] V L Jaya and R Gopikakumari ldquoIEM a new image enhance-ment metric for contrast and sharpness measurementsrdquo Inter-national Journal of Computer Applications vol 79 no 9 pp 1ndash92013
[35] S S Agaian K Panetta and A M Grigoryan ldquoTransform-based image enhancement algorithms with performance mea-surerdquo IEEE Transactions on Image Processing vol 10 no 3 pp367ndash382 2001
[36] M Braiki A Benzinou K Nasreddine S Labidi and NHymery ldquoSegmentation of dendritic cells from microscopicimages using mathematical morphologyrdquo in Proceedings ofthe 2nd International Conference on Advanced Technologies forSignal and Image Processing (ATSIP rsquo16) pp 282ndash287 MonastirTunisia March 2016
[37] L R Dice ldquoMeasures of the amount of ecologic associationbetween speciesrdquo Ecology vol 26 no 3 pp 297ndash302 1945
[38] P Jaccard ldquoThe distribution of the flora in the alpine zonerdquoNewPhytologist vol 11 no 2 pp 37ndash50 1912
images in which the cell regions are relatively obvious to dotest Figure 6 demonstrates one group segmentation resultwhere Figure 6(a) is the original RGB pathological imageFigure 6(b) is the ground truth segmented manually underthe guidance of pathologists and Figure 6(c)ndashFigure 6(j)are obtained from segmenting the corresponding enhancedimages
It is clear that the cell segmentation result of ouralgorithm is more close to the ground truth In order tocompare the segmentation results quantitatively we adopttwo standard segmentation metrics to evaluate namely dicecoefficient [37] and Jaccard index [38] The definitions ofthem are as follows
DC (SRTR) = 2 timesNum (pixelSR cap pixelTR)Num (pixelSR) +Num (pixelTR)
JI (SRTR) = Num (pixelSR cap pixelTR)Num (pixelSR cup pixelTR)
(15)
where SR is the segmentation region TR is the true regionof the target and Num(pixel) is the number of relevantpixel points The average results of performance metricsare displayed in Table 4 We can see that the result ofour proposed method is the best Our HDR pathologicalimage enhancement method is superior to the comparisonalgorithms as we can improve the image segmentation andpathological analysis better
In sum we proposed that HDR pathological imageenhancement method obtains a better result according to
both pathologistsrsquo subjective evaluation and quantitativeanalysis in data and the cell segmentationmethod also provesthe better performance of our method as the quality anddetail of original image are both improved
5 Conclusions
This paper proposes new HDR pathological image enhance-ment methods based on GIF and improved bias fieldcorrection model First stain normalization and waveletdenoising operations are used in image preprocessing Andthe improved bias field model is introduced to correct theintensity inhomogeneity and detail discontinuity of imageThen the HDR pathological image is generated using LDRimage and H and E channel images Next the 119884 componentof HDR image is separated into base layer and detail layerby GIF and the two layers are enhanced separately Finallythe fine enhanced image is acquired after combining the 119884component and the color components To verify the effective-ness of the proposed method we perform the enhancementexperiments using 140 pathological images The experimentresults and comparisons with related work demonstrate thatour proposed method improves the image quality in terms ofhuman vision PSNR SD mean EME information entropyand cell segmentation
Competing Interests
The authors declare that there were no competing interestsregarding the publication of this article
10 BioMed Research International
Acknowledgments
This work was supported by the National Natural ScienceFoundation of China (no 61472073)
References
[1] SM Pizer E P Amburn J D Austin et al ldquoAdaptive histogramequalization and its variationsrdquo Computer Vision Graphics ampImage Processing vol 39 no 3 pp 355ndash368 1987
[2] V Vani and K V M Prashanth ldquoColor image enhancementtechniques in Wireless Capsule Endoscopyrdquo in Proceedings ofthe IEEE International Conference on Trends in AutomationCommunications and Computing Technology (I-TACT rsquo15) vol1 pp 1ndash6 Bangalore India December 2015
[3] H Cao L Tian J Liu H Wang and S Feng ldquoColor imageenhancement using power-constraint histogram equalizationfor AMOLEDrdquo in Proceedings of the IEEE 11th InternationalConference on ASIC (ASICON rsquo15) pp 1ndash4 IEEE ChengduChina November 2015
[4] N M Kwok G Fang and H Y Shi ldquoColor enhancementfor images from digital camera using a transformation-freeapproachrdquo in Proceedings of the 9th International Conferenceon Sensing Technology (ICST rsquo15) pp 168ndash172 IEEE AucklandNew Zealand December 2015
[5] S D Nikam and R U Yawale ldquoColor image enhancementusing daubechies wavelet transform and HIS color modelrdquoin Proceedings of the International Conference on IndustrialInstrumentation and Control (ICIC rsquo15) pp 1323ndash1327 IEEEPune India May 2015
[6] L G Villanueva G M Callico F Tobajas et al ldquoMedicaldiagnosis improvement through image quality enhancementbased on super-resolutionrdquo in Proceedings of the 13th EuromicroConference onDigital SystemDesign ArchitecturesMethods andTools (DSD rsquo10) pp 259ndash262 IEEE Lille France September2010
[7] W Sun F Li and Q Zhang ldquoThe applications of improvedretinex algorithm for X-ray medical image enhancementrdquo inProceedings of the International Conference on Computer Scienceand Service System (CSSS rsquo12) pp 1655ndash1658 IEEE NanjingChina August 2012
[8] G Zhang D Sun P Yan H Zhao and Z Li ldquoA LDCT imagecontrast enhancement algorithm based on single-scale retinextheoryrdquo in Proceedings of the International Conference on Com-putational Intelligence for Modelling Control amp Automation pp1282ndash1287 IEEE Computer Society Vienna Austria December2008
[9] S Setty N K Srinath and M C Hanumantharaju ldquoDevel-opment of multiscale retinex algorithm for medical imageenhancement based on multi-rate samplingrdquo in Proceedingsof the International Conference on Signal Processing ImageProcessing amp Pattern Recognition pp 145ndash150 2013
[10] J Mccann ldquoLessons learned from mondrians applied to realimages and color gamutsrdquo in Proceedings of the Color andImaging Conference vol 8 pp 1ndash8 1999
[11] B V Funt F Ciurea and J J McCann ldquoRetinex in Matlabrdquo inProceedings of the Color and Imaging Conference pp 112ndash121Scottsdale Ariz USA November 2000
[12] K Kim J Bae and J Kim ldquoNatural hdr image tone mappingbased on retinexrdquo IEEE Transactions on Consumer Electronicsvol 57 no 4 pp 1807ndash1814 2011
[13] J Kuang G M Johnson and M D Fairchild ldquoiCAM06 arefined image appearance model for HDR image renderingrdquoJournal of Visual Communication amp Image Representation vol18 no 5 pp 406ndash414 2007
[14] M-L Song H-Q Wang C Chen X-Q Ye and W-K GuldquoTone mapping for high dynamic range image using a proba-bilistic modelrdquo Journal of Software vol 20 no 3 pp 734ndash7432010
[15] F Branchitta M Diani G Corsini and M Romagnoli ldquoNewtechnique for the visualization of high dynamic range infraredimagesrdquo Optical Engineering vol 48 no 9 Article ID 0964012009
[16] C Zuo ldquoDisplay and detail enhancement for high-dynamic-range infrared imagesrdquo Optical Engineering vol 50 no 12Article ID 127401 pp 895ndash900 2011
[17] K He J Sun and X Tang ldquoGuided image filteringrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol35 no 6 pp 1397ndash1409 2013
[18] N Liu and D Zhao ldquoDetail enhancement for high-dynamic-range infrared images based on guided image filterrdquo InfraredPhysics amp Technology vol 67 pp 138ndash147 2014
[19] C Li R Huang Z Ding et al ldquoA level set method for imagesegmentation in the presence of intensity inhomogeneities withapplication toMRIrdquo IEEE Transactions on Image Processing vol20 no 7 pp 2007ndash2016 2011
[20] C Li J C Gore and C Davatzikos ldquoMultiplicative intrinsiccomponent optimization (MICO) for MRI bias field estimationand tissue segmentationrdquoMagnetic Resonance Imaging vol 32no 7 pp 913ndash923 2014
[21] A Vahadane T Peng A Sethi et al ldquoStructure-preservingcolor normalization and sparse stain separation for histologicalimagesrdquo IEEE Transactions on Medical Imaging vol 35 no 8pp 1962ndash1971 2016
[22] E Reinhard M Ashikhmin B Gooch and P Shirley ldquoColortransfer between imagesrdquo IEEE Computer Graphics amp Applica-tions vol 21 no 5 pp 34ndash41 2001
[23] D L Donoho ldquoDe-noising by soft-thresholdingrdquo IEEE Trans-actions on Information Theory vol 41 no 3 pp 613ndash627 1995
[24] E Zhang H Yang and M Xu ldquoA novel tone mappingmethod for high dynamic range image by incorporating edge-preserving filter into method based on retinexrdquo Applied Mathe-matics amp Information Sciences vol 9 no 1 pp 411ndash417 2015
[25] M Macenko M Niethammer J S Marron et al ldquoA methodfor normalizing histology slides for quantitative analysisrdquo inProceedings of the IEEE International Conference on Symposiumon Biomedical Imaging From Nano To Macro pp 1107ndash1110IEEE Press 2009
[26] T Mitsunaga and S K Nayar ldquoRadiometric self calibrationrdquoin Proceedings of the IEEE Computer Society Conference onComputer Vision and Pattern Recognition vol 1 p 1374 FortCollins Colo USA June 1999
[27] F Durand and J Dorsey ldquoFast bilateral filtering for the displayof high-dynamic-range imagesrdquoACMTransactions onGraphicsvol 21 no 3 pp 257ndash266 2002
[28] D J Jobson Z-U Rahman andG AWoodell ldquoProperties andperformance of a centersurround retinexrdquo IEEE Transactionson Image Processing vol 6 no 3 pp 451ndash462 1997
[29] D J Jobson Z-U Rahman and G A Woodell ldquoA multiscaleretinex for bridging the gap between color images and thehuman observation of scenesrdquo IEEE Transactions on ImageProcessing vol 6 no 7 pp 965ndash976 1997
BioMed Research International 11
[30] YWangQ Chen andB Zhang ldquoImage enhancement based onequal area dualistic sub-image histogram equalizationmethodrdquoIEEE Transactions on Consumer Electronics vol 45 no 1 pp68ndash75 1999
[31] S-D Chen and A R Ramli ldquoMinimum mean brightnesserror bi-histogram equalization in contrast enhancementrdquo IEEETransactions on Consumer Electronics vol 49 no 4 pp 1310ndash1319 2003
[32] S-D Chen and A R Ramli ldquoContrast enhancement usingrecursive mean-separate histogram equalization for scalablebrightness preservationrdquo IEEE Transactions on Consumer Elec-tronics vol 49 no 4 pp 1301ndash1309 2003
[33] K S Sim C P Tso and Y Y Tan ldquoRecursive sub-imagehistogram equalization applied to gray scale imagesrdquo PatternRecognition Letters vol 28 no 10 pp 1209ndash1221 2007
[34] V L Jaya and R Gopikakumari ldquoIEM a new image enhance-ment metric for contrast and sharpness measurementsrdquo Inter-national Journal of Computer Applications vol 79 no 9 pp 1ndash92013
[35] S S Agaian K Panetta and A M Grigoryan ldquoTransform-based image enhancement algorithms with performance mea-surerdquo IEEE Transactions on Image Processing vol 10 no 3 pp367ndash382 2001
[36] M Braiki A Benzinou K Nasreddine S Labidi and NHymery ldquoSegmentation of dendritic cells from microscopicimages using mathematical morphologyrdquo in Proceedings ofthe 2nd International Conference on Advanced Technologies forSignal and Image Processing (ATSIP rsquo16) pp 282ndash287 MonastirTunisia March 2016
[37] L R Dice ldquoMeasures of the amount of ecologic associationbetween speciesrdquo Ecology vol 26 no 3 pp 297ndash302 1945
[38] P Jaccard ldquoThe distribution of the flora in the alpine zonerdquoNewPhytologist vol 11 no 2 pp 37ndash50 1912
This work was supported by the National Natural ScienceFoundation of China (no 61472073)
References
[1] SM Pizer E P Amburn J D Austin et al ldquoAdaptive histogramequalization and its variationsrdquo Computer Vision Graphics ampImage Processing vol 39 no 3 pp 355ndash368 1987
[2] V Vani and K V M Prashanth ldquoColor image enhancementtechniques in Wireless Capsule Endoscopyrdquo in Proceedings ofthe IEEE International Conference on Trends in AutomationCommunications and Computing Technology (I-TACT rsquo15) vol1 pp 1ndash6 Bangalore India December 2015
[3] H Cao L Tian J Liu H Wang and S Feng ldquoColor imageenhancement using power-constraint histogram equalizationfor AMOLEDrdquo in Proceedings of the IEEE 11th InternationalConference on ASIC (ASICON rsquo15) pp 1ndash4 IEEE ChengduChina November 2015
[4] N M Kwok G Fang and H Y Shi ldquoColor enhancementfor images from digital camera using a transformation-freeapproachrdquo in Proceedings of the 9th International Conferenceon Sensing Technology (ICST rsquo15) pp 168ndash172 IEEE AucklandNew Zealand December 2015
[5] S D Nikam and R U Yawale ldquoColor image enhancementusing daubechies wavelet transform and HIS color modelrdquoin Proceedings of the International Conference on IndustrialInstrumentation and Control (ICIC rsquo15) pp 1323ndash1327 IEEEPune India May 2015
[6] L G Villanueva G M Callico F Tobajas et al ldquoMedicaldiagnosis improvement through image quality enhancementbased on super-resolutionrdquo in Proceedings of the 13th EuromicroConference onDigital SystemDesign ArchitecturesMethods andTools (DSD rsquo10) pp 259ndash262 IEEE Lille France September2010
[7] W Sun F Li and Q Zhang ldquoThe applications of improvedretinex algorithm for X-ray medical image enhancementrdquo inProceedings of the International Conference on Computer Scienceand Service System (CSSS rsquo12) pp 1655ndash1658 IEEE NanjingChina August 2012
[8] G Zhang D Sun P Yan H Zhao and Z Li ldquoA LDCT imagecontrast enhancement algorithm based on single-scale retinextheoryrdquo in Proceedings of the International Conference on Com-putational Intelligence for Modelling Control amp Automation pp1282ndash1287 IEEE Computer Society Vienna Austria December2008
[9] S Setty N K Srinath and M C Hanumantharaju ldquoDevel-opment of multiscale retinex algorithm for medical imageenhancement based on multi-rate samplingrdquo in Proceedingsof the International Conference on Signal Processing ImageProcessing amp Pattern Recognition pp 145ndash150 2013
[10] J Mccann ldquoLessons learned from mondrians applied to realimages and color gamutsrdquo in Proceedings of the Color andImaging Conference vol 8 pp 1ndash8 1999
[11] B V Funt F Ciurea and J J McCann ldquoRetinex in Matlabrdquo inProceedings of the Color and Imaging Conference pp 112ndash121Scottsdale Ariz USA November 2000
[12] K Kim J Bae and J Kim ldquoNatural hdr image tone mappingbased on retinexrdquo IEEE Transactions on Consumer Electronicsvol 57 no 4 pp 1807ndash1814 2011
[13] J Kuang G M Johnson and M D Fairchild ldquoiCAM06 arefined image appearance model for HDR image renderingrdquoJournal of Visual Communication amp Image Representation vol18 no 5 pp 406ndash414 2007
[14] M-L Song H-Q Wang C Chen X-Q Ye and W-K GuldquoTone mapping for high dynamic range image using a proba-bilistic modelrdquo Journal of Software vol 20 no 3 pp 734ndash7432010
[15] F Branchitta M Diani G Corsini and M Romagnoli ldquoNewtechnique for the visualization of high dynamic range infraredimagesrdquo Optical Engineering vol 48 no 9 Article ID 0964012009
[16] C Zuo ldquoDisplay and detail enhancement for high-dynamic-range infrared imagesrdquo Optical Engineering vol 50 no 12Article ID 127401 pp 895ndash900 2011
[17] K He J Sun and X Tang ldquoGuided image filteringrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol35 no 6 pp 1397ndash1409 2013
[18] N Liu and D Zhao ldquoDetail enhancement for high-dynamic-range infrared images based on guided image filterrdquo InfraredPhysics amp Technology vol 67 pp 138ndash147 2014
[19] C Li R Huang Z Ding et al ldquoA level set method for imagesegmentation in the presence of intensity inhomogeneities withapplication toMRIrdquo IEEE Transactions on Image Processing vol20 no 7 pp 2007ndash2016 2011
[20] C Li J C Gore and C Davatzikos ldquoMultiplicative intrinsiccomponent optimization (MICO) for MRI bias field estimationand tissue segmentationrdquoMagnetic Resonance Imaging vol 32no 7 pp 913ndash923 2014
[21] A Vahadane T Peng A Sethi et al ldquoStructure-preservingcolor normalization and sparse stain separation for histologicalimagesrdquo IEEE Transactions on Medical Imaging vol 35 no 8pp 1962ndash1971 2016
[22] E Reinhard M Ashikhmin B Gooch and P Shirley ldquoColortransfer between imagesrdquo IEEE Computer Graphics amp Applica-tions vol 21 no 5 pp 34ndash41 2001
[23] D L Donoho ldquoDe-noising by soft-thresholdingrdquo IEEE Trans-actions on Information Theory vol 41 no 3 pp 613ndash627 1995
[24] E Zhang H Yang and M Xu ldquoA novel tone mappingmethod for high dynamic range image by incorporating edge-preserving filter into method based on retinexrdquo Applied Mathe-matics amp Information Sciences vol 9 no 1 pp 411ndash417 2015
[25] M Macenko M Niethammer J S Marron et al ldquoA methodfor normalizing histology slides for quantitative analysisrdquo inProceedings of the IEEE International Conference on Symposiumon Biomedical Imaging From Nano To Macro pp 1107ndash1110IEEE Press 2009
[26] T Mitsunaga and S K Nayar ldquoRadiometric self calibrationrdquoin Proceedings of the IEEE Computer Society Conference onComputer Vision and Pattern Recognition vol 1 p 1374 FortCollins Colo USA June 1999
[27] F Durand and J Dorsey ldquoFast bilateral filtering for the displayof high-dynamic-range imagesrdquoACMTransactions onGraphicsvol 21 no 3 pp 257ndash266 2002
[28] D J Jobson Z-U Rahman andG AWoodell ldquoProperties andperformance of a centersurround retinexrdquo IEEE Transactionson Image Processing vol 6 no 3 pp 451ndash462 1997
[29] D J Jobson Z-U Rahman and G A Woodell ldquoA multiscaleretinex for bridging the gap between color images and thehuman observation of scenesrdquo IEEE Transactions on ImageProcessing vol 6 no 7 pp 965ndash976 1997
BioMed Research International 11
[30] YWangQ Chen andB Zhang ldquoImage enhancement based onequal area dualistic sub-image histogram equalizationmethodrdquoIEEE Transactions on Consumer Electronics vol 45 no 1 pp68ndash75 1999
[31] S-D Chen and A R Ramli ldquoMinimum mean brightnesserror bi-histogram equalization in contrast enhancementrdquo IEEETransactions on Consumer Electronics vol 49 no 4 pp 1310ndash1319 2003
[32] S-D Chen and A R Ramli ldquoContrast enhancement usingrecursive mean-separate histogram equalization for scalablebrightness preservationrdquo IEEE Transactions on Consumer Elec-tronics vol 49 no 4 pp 1301ndash1309 2003
[33] K S Sim C P Tso and Y Y Tan ldquoRecursive sub-imagehistogram equalization applied to gray scale imagesrdquo PatternRecognition Letters vol 28 no 10 pp 1209ndash1221 2007
[34] V L Jaya and R Gopikakumari ldquoIEM a new image enhance-ment metric for contrast and sharpness measurementsrdquo Inter-national Journal of Computer Applications vol 79 no 9 pp 1ndash92013
[35] S S Agaian K Panetta and A M Grigoryan ldquoTransform-based image enhancement algorithms with performance mea-surerdquo IEEE Transactions on Image Processing vol 10 no 3 pp367ndash382 2001
[36] M Braiki A Benzinou K Nasreddine S Labidi and NHymery ldquoSegmentation of dendritic cells from microscopicimages using mathematical morphologyrdquo in Proceedings ofthe 2nd International Conference on Advanced Technologies forSignal and Image Processing (ATSIP rsquo16) pp 282ndash287 MonastirTunisia March 2016
[37] L R Dice ldquoMeasures of the amount of ecologic associationbetween speciesrdquo Ecology vol 26 no 3 pp 297ndash302 1945
[38] P Jaccard ldquoThe distribution of the flora in the alpine zonerdquoNewPhytologist vol 11 no 2 pp 37ndash50 1912
[30] YWangQ Chen andB Zhang ldquoImage enhancement based onequal area dualistic sub-image histogram equalizationmethodrdquoIEEE Transactions on Consumer Electronics vol 45 no 1 pp68ndash75 1999
[31] S-D Chen and A R Ramli ldquoMinimum mean brightnesserror bi-histogram equalization in contrast enhancementrdquo IEEETransactions on Consumer Electronics vol 49 no 4 pp 1310ndash1319 2003
[32] S-D Chen and A R Ramli ldquoContrast enhancement usingrecursive mean-separate histogram equalization for scalablebrightness preservationrdquo IEEE Transactions on Consumer Elec-tronics vol 49 no 4 pp 1301ndash1309 2003
[33] K S Sim C P Tso and Y Y Tan ldquoRecursive sub-imagehistogram equalization applied to gray scale imagesrdquo PatternRecognition Letters vol 28 no 10 pp 1209ndash1221 2007
[34] V L Jaya and R Gopikakumari ldquoIEM a new image enhance-ment metric for contrast and sharpness measurementsrdquo Inter-national Journal of Computer Applications vol 79 no 9 pp 1ndash92013
[35] S S Agaian K Panetta and A M Grigoryan ldquoTransform-based image enhancement algorithms with performance mea-surerdquo IEEE Transactions on Image Processing vol 10 no 3 pp367ndash382 2001
[36] M Braiki A Benzinou K Nasreddine S Labidi and NHymery ldquoSegmentation of dendritic cells from microscopicimages using mathematical morphologyrdquo in Proceedings ofthe 2nd International Conference on Advanced Technologies forSignal and Image Processing (ATSIP rsquo16) pp 282ndash287 MonastirTunisia March 2016
[37] L R Dice ldquoMeasures of the amount of ecologic associationbetween speciesrdquo Ecology vol 26 no 3 pp 297ndash302 1945
[38] P Jaccard ldquoThe distribution of the flora in the alpine zonerdquoNewPhytologist vol 11 no 2 pp 37ndash50 1912
