Digital Mammography: Comparison of Adaptive and Nonadaptive
CAD Methods for Mass Detection 1
We Qian, PhD, Lihua Li, PhD, Laurence Clarke, PhD Robert A. Clark, MD, Jerry Thomas, CAPT, MSC, USN
Rationale and Objectives. The authors compared the performance of adaptive and nonadaptive computer-aided diagnostic (CAD) methods for breast mass detection with digital mammography.
Materials and Methods. Both adaptive and nonadaptive modular CAD methods employed recent advances in multiresolution and mutiorientation wavetet transforms for improved feature extraction. The nonadaptive method uses fixed parameters for the image preprocessing mod- ules. The adaptive method, a new class of algorithms, adapts to image content by selecting parameters for the image preprocessing modules within a parameter range. Comparison of the two methods was performed for each individual CAD module with a region-of-interest (ROD database containing all mass types and normal tissue.
Results. Receiver operating characteristic (ROC) analysis clearly demonstrated an improvement in performance for the three adaptive modules and a significant overall dif- ference between the two methods. The average ROC area
, index (Az) values were 0.86 and 0.95 for the nonadaptive and adaptive methods, respectively. The corresponding P value is .0145. For a previously reported database of full manunographic images containing 50 abnormal cases with all mass types and 50 normal images, the adaptive CAD method had a sensitivity of 96% (1.71 false-positive results per image) compared with 89% (1.91 false-posi- tive results per image) for the nonadaptive CAD method.
Conclusion. The adaptive CAD method demonstrated better performance. A study is in progress to determine the generalizability of the adaptive CAD method by ap- plying it to larger retrospective image databases with dif- ferent film digitizers.
Key Words. Breast, diseases; breast radiography, tech- nology; computers, diagnostic aid.
Computer-assisted diagnostic (CAD) methods have been proposed as a second-opinion strategy for breast cancer screening with digital mammography. These methods have
focused on the detection of either microcalcification clus-
ters or masses (1). Mass detection poses a more difficult problem because masses often have varying sizes, shapes,
and densities; exhibit poor image contrast, particularly for
dense breasts; are highly connected to the surrounding pa- renchymal tissue, particularly in the case of spiculated le- sions; and are surrounded by a nonuniform tissue back- ground with similar characteristics (2-4). The segmenta-
tion of masses and the computation of the related pixel intensity and morphologic and directional texture features require improved feature-extraction methods to distinguish masses from normal tissues. Examples of such features in-
clude mass shape, mass margins, or spiculations (5). Mass detection is proposed here as a good clinical model for the evaluation of a new class of adaptive CAD methods for
image preprocessing, designed to improve feature extrac- tion. The adaptive methods proposed are also useful for other CAD applications, such as the detection of micro- calcifications and lung nodules.
Image-preprocessing methods have previously been
proposed for mass segmentation and extraction of related morphologic features. They have included the use of histo- gram equalization and thresholding of texture energy im-
ages (6) or density-weighted filtering of image contrast
Acad Radiol 1999; 6:471-480
1 From the Department of Radiology, College of Medicine, and the H. Lee Moffitt Cancer Center and Research Institute, University of South Florida, 12901 Bruce B. Downs BIvd, Box 17, Tampa, FL 33612-4799 (W.Q., L.L., L.C., R.A.C.), and the Uniformed Services University of the Health Sci- ences, Bethesda, Md (J.T0. Received June 15, 1998; revision requested November 19; revision received February 12, 1999; accepted March 1. Ad- dress reprint requests to W.Q.
©AUR, 1999
471
enhancement (7). These image enhancement methods are generally single-scale approaches for each image-prepro-
cessing CAD module. The use of multiresolution ap- proaches to image segmentation, based on fuzzy pyramid
linking (8), has been proposed, as these methods should potentially improve image segmentation and feature ex-
traction. To date, multiresolution methods have not been
proposed in which the CAD parameters for image prepro- cessing are selected adaptively on a per-image basis,
rather than by the use of fixed parameters. Feature-extraction methods have been used for the
detection of spiculated lesions. These methods include (a) the use of Laws texture analysis of edge orientation to identify areas of locally radiating structure with false- positive reduction (5-9) and (b) the use of gray-scale seed growing to measure the angular distribution of ra-
dial gradient histograms for the detection of spiculations within a defined region of interest (ROI), which requires manual intervention (10). A more direct and automatic
measure of the presence of spiculations has been reported
that includes radial line enhancement followed by a Bayesian combination of cues generated by the Hough transform (11) and the detection of stellate patterns (without assuming the presence of a central mass with line orientation) by three second-order Gaussian deriva- tive operators (12,13). The angular distribution of these methods was approximately of the order of four to eight
orientations over a 360 ° rotation. The direct measurement
of the presence of spiculations and the subtraction of di- rectional features to improve detection of suspicious ar- eas, previously reported by these investigators (14), are important steps for improving CAD methods. Where higher directional orientations are present in the mammo- graphic image, there is a need to improve directionality
of directional wavlet transforms and to include the use of adaptive methods.
This article reports the development and evaluation of
a new class of adaptive CAD algorithms for the detection of masses in digitized mammograms. The adaptive meth- ods are an extension of a previously reported nonadaptive method (15), with which they are compared. Both the
nonadaptive and adaptive CAD methods employ three
key image-preprocessing CAD modules specifically tai- lored for digital mammography, namely, (a) a tree-struc- tured nonlinear filter for noise suppression, (b) a multi- orientation directional wavelet transform for the removal
of directional features and for direct detection of spicula- tions, and (c) a multiresolution wavelet transform for im- age enhancement to improve the segmentation of suspi-
cious areas. This modular approach to design allows the
performance of each CAD module to be independently
evaluated, which is critical for CAD algorithm optimiza- tion. The preservation of image detail, such as morpho-
logic and directional features, is also emphasized to allow the extension of these methods for automatic mass classi-
fication. Finally, the use of adaptive image preprocessing
is an important step in the development of a more gener- alized CAD method that is independent of the digital sen-
sor and thus suitable for multicenter clinical trials. The purpose of this study was to compare the perfor-
mance of adaptive and nonadaptive CAD methods for breast mass detection with digital mammography. Receiver operating characteristic (ROC) analysis and computer free- response ROC curves are used to evaluate the performance of each module and of the overall CAD system. The study
used two image databases for which electronic ground truth (ie, truth file) had been established. One database, based on ROI data, included both masses and normal tissue to allow
relative comparison of individual CAD modules. The sec- ond database included full mammographic images to allow a relative comparison of the overall CAD methods. Some previously reported case databases were used, in addition to the two previous databases, to minimize problems related to case selection in terms of performance evaluation.
CAD Modules
Figure 1 shows a schematic diagram of the adaptive and nonadaptive CAD modules for mass detection de- scribed in this section. The theoretical details of several of the individual CAD modules have been previously re- ported (15-18).
Tree-structured nonlinear filtering CAD module . - This module is a three-stage filter designed with centrally
weighted median filters as subfiltering blocks (17,19). These blocks were applied to each pixel within the filter window (18). To preserve image detail (such as paren- chymal tissue structure), modified windows were used for the filter bank in the first stage and the filtered image
was compared to the raw image in each stage. This filter- ing method significantly improved background noise re-
duction compared with that of single-stage filters. Noise reduction was determined by using simulated images
with different noise levels and representative mammo-
graphic images (17,19-20). The filtering method was used in this study for improved noise suppression prior to the use of wavelet transforms.
472
Vol 6, No 8, August 1999 ADAPTIVE AND NONADAPTIVE CAD METHODS
Digitized image ) 220 / 180 gm
Noise suppression and artifact removal Tree-structured filter (TSF)
Gray scale domain
Multiorientation or directional wavelet transform (DWT) decomposition(N=8)
• Directional feature image • Differentiation of spiculations/
directional features by ray tracing
Texture domain I
• Smoothed Image.
• Multiresolution enhancement M= 2 TSWT Adaetive
I ~v~ ,,.,..,~ I t[,,,,"'kv~vl [± I " ; ~t j ] 4 kvL;-1 ~,lvl ~ I ~ - -
Morphological Domain I
Feature extraction: All domains ]
I Feature classification Fuzzy binary decision tree (FBDT)
I I Figure 1. Block diagram of proposed CAD modules for mass detection. The directional wavelet transform (DW-I) CAD module provides directional feature images and differentiation of directional features and spiculations and a smoothed image for improved image segmen- tation by multiresolution two-channel tree-structured wavelet transform (TSW'r) and adaptive clustering CAD modules. The image-pre- processing methods allow improved feature extraction in all three domains, as shown in the shaded areas. Adaptive CAD modules for mass detection include the adaptive tree-structured nonlinear filter (TSF) module, directional wavelet transform, and four-channel tree- structured wavelet transform shown in the shaded areas. FROC = free-response ROC.
The advantage of the current tree-structured nonlinear filter module for image noise suppression is that its appli-
cation does not require a priori knowledge of the local sta-
tistics within the filter window--that is, this method is nonadaptive and therefore computationally efficient (18).
Although this module has already demonstrated good per- formance, we attempted to optimize it by the application of
adaptive methods. The adaptive criteria included the devel- opment of a technique for automatic parameter selection
for the module (ie, parameters K l, K s, and K 3, as defined in
473
Eqq.[7]-[10] in [18]) and the development of a method for selecting filter window sizes (ie, from 3 x 3 to 7 x 7),
depending on requirements for image detail preservation. The initial physical performance of the adaptive filter was evaluated based on standard signal processing criteria: (a) evaluation of localized metrics for noise, such as nor-
malized mean square error and difference images to show
structured noise, (b) evaluation of images simulating the American College of Radiology (ACR) test phantom with
different levels of Poisson and signal-dependent noise (18), and (c) evaluation of possible artifact generation by
application of the directional wavelet transform module to the same simulated images.
Directional wavelet transform CAD module.--This module, a wavelet transform for multiorientation signal
decomposition, was implemented by using polar coor- dinates (as opposed to rectangular coordinates) on a pixel-
by-pixel basis (21). Selection of the wavelet basis func- tions was important for high-orientation selectivity and
was achieved by introducing a directional sensitivity con- straint on the wavelet function (15). The input image was decomposed by the directional wavelet transform, yielding two output images. The first was a directional texture im- age, used for directional feature extraction. The second was a smoothed version of the original image, with direc- tional information removed, and was later used for image segmentation, as illustrated in Figure 1. This CAD module had two important roles: isolation of the central mass from
the surrounding parenchymal tissue for segmentation of stellate tumors and direct detection and analysis of spicula- tions and their differentiation from other directional fea- tures in the mammogram with a ray-tracing algorithm. The previous form had a fixed order of directionality of n = 8 orientations.
The adaptive version of this module was designed as a
bank of wavelet filters implemented by the use of adaptive combiners with different weight factors (15), which al- lowed it to be uniquely modified for a higher order n-di- rectional filter. For example, a higher-order wavelet orien- tation (n = 16) was recently implemented, affecting the di- rection angle Qi--that is, the directional bandwidth of the
wavelet functions--for more selective extraction of direc- tional features. Further improvement in robustness was
achieved through adaptivity of the weights W i applied to the directional features (14,15). We performed an initial evaluation of n = 16 and found an improvement in the
preservation of the shape of segmented suspicious areas and in the identification of spiculation for selected spicu- lated cases (14). However, for a spiculated mass, the angu-
lar distribution of spiculations or the mass margins may vary around its boundary. For the adaptive criteria, n was
selected for each pixel point to match these changes. By adaptively selecting n, sampling was potentially improved for the detection of speculations that might lie inside more than one arc. Alternatively, for higher orders of n, an im-
provement in the signal-to-noise ratio for the detection of a spiculation was possible. The range of n, which influ-
ences the angular bandwidth frequency and directional
sensitivity, was adaptively selected within 4-32, which corresponds to a 45°-5.63 ° arc width. As described previ-
ously (14), this arc width provided higher sampling and a more mathematically rigorous method. The physical per- formance of the adaptive method was initially evaluated with the same simulated ACR image, which contained lin-
ear and other structures, to determine whether appropriate structures were identified or any artifacts generated.
Image enhancement CAD module.--A two-channel tree-structured wavelet transform (as opposed to a single-
scale, region-based image enhancement method [22]) was previously proposed as an efficient multiresolution enhancement method for suspicious areas (22). The lower-resolution characteristics were useful for localiza- tion of the suspicious areas, while the higher resolution information was essential for fine detail, such as mass- margin feature extraction. In earlier work, we showed the importance of image enhancement with two-channel tree- structured and directional wavelet transforms, prior to
segmentation of suspicious areas, by comparing the per- formance of the CAD module with and without wavelet transforms, which improved extraction of the morpho- logic features (17).
In the adaptive enhancement, subimages were selected with localized metrics. These metrics included a higher-or- der multichannel correlation function and an energy func-
tion to identify dominant subimages (dominant frequency and resolution channels) (23). This approach allowed for variations in image contrast observed over a large data- base. In the adaptive method, a four-channel tree-struc- tured wavelet transform was used (Fig 1) with an M x M different resolution information, generating M 2 or 16
subimages (16,24). The use of a four-channel wavelet transform increased the choice for selection of the 16 subimages, which should improve the enhancement of sus-
picious areas and their subsequent segmentation (24). Adaptive selection of the subimages by using localized
metrics to zoom onto the desired features for mass en- hancement was then possible by training a neural network with a small training database that contained images of
474
Vol 6, No 8, August 1999 ADAPTIVE AND NONADAPTIVE CAD METHODS
varying breast density. For localized metrics, an energy function and a correlation function had been previously
used. Another four adaptive criteria will be explored in this study--pixel-intensity variation, mean intensity difference, mean gradient of region boundary, and circularity of suspi- cious areas. The former two metrics will be used for global screening of the full mammogram, by means of a 64 x 64-
pixel window with 50% overlap, with the larger values ini-
tially identifying suspicious areas. The latter four metrics will then be automatically computed from the suspicious ar- eas. The total of six metrics will then be used to determine
the optimum number of subimages for enhancement of sus- picious areas. For this purpose, the neural network will be used as the classifier, and the cross-validation method will be used to train the neural network. The adaptive clustering method performed well in the preliminary data in this study
and will be unchanged. The initial evaluation of this CAD segmentation module will involve analysis of the seg- mented images of the mass versus the ground truth provided
by the full image database.
Adaptive clustering CAD module. The adaptive cluster- ing method used was the same for both adaptive and non- adaptive CAD methods. It was based on an adaptive con- strained distance function that had more robust performance
than the nonadaptive module, as previously reported (25). Segmentation of the suspicious areas into two classes is per- formed (25) for the full mammogram (fatty tissue and pa- renchymal tissue or suspicious masses)--specifically for
the wavelet-enhanced image (Fig 1). Note that traditional clustering methods are based on a distance function crite- rion (11) that may not optimally differentiate background and suspicious areas and may be more image dependent.
Feature extraction was then performed in the gray-level
domain, morphologic domain, and directional feature do- main. Three features in gray-level domain included inten- sity variation, which measures the smoothness of the pixel intensity in the extracted region; mean intensity difference,
which measures the intensity difference between the ex- tracted region and its surrounding areas; and mean gradient of region boundary, which measures the edge contrast of the extracted region. Three features in the morphologic do-
main were areas, which includes the number of pixels in the extracted region; circularity, which provides information
about gross shape of the mass but contributes nothing spe- cific about the fine details of the mass boundary; and nor-
malized deviation of radial length, whic h is a measure of
how the boundary changes in a microscopic way. Three fea- tures in the directional domain included the number of spiculations, the orientation of the spiculations, and the av-
erage normalized spiculation length. Directional feature ex-
traction included contour tracing of the segmented region
and extraction of the spiculations for stellate lesions. A ray- tracing algorithm was developed to determine the direc-
tional features. The ray-tracing algorithm included three steps: (a) detecting the mass contours based on the seg-
mented images, in which the directional information was re- moved (see example in Fig 2b); (b) eliminating background
detail signals, including weak random directional lines, and
isolating the strong directional signals (possible spiculations based on directional feature images shown in Fig 2c, 2d), and (c) combining the images obtained from the preceding steps, removing all signals inside the mass contour, and
tracing the spiculation signals, starting from the mass con- tour, by linking and searching the spiculation lines.
As described elsewhere (15,26), the investigation used this classification method rather than neural network ap-
proaches because the former allowed additional features to be added more easily. The module was modified by adding
two more decision nodes based on the spiculation features
at levels 5 and 6 to the output of the decision nodes, to en- hance the detection of stellate masses (15,26).
I m a g e Da tabases ROI database.--To compare the relative performance
of individual CAD modules for optimization, four sets of
50 ROIs from mammographic images were selected. Each ROI was composed of 256 x 256 pixels at 16 bits. The
four sets included (a) the areas containing biopsy-proved masses for all mass types (20 stellate, 14 circumscribed, and 16 irregular masses, with 16 cases of minimal cancers [< 1 cm]), identified by experienced radiologists, to allow a relative comparison of CAD methods (27), (b) the re-
gions containing the dense tissue, (c) the regions of mixed tissues, and (d) the regions of fatty tissues. The four sets of ROIs, each with 50 samples, were randomly and equally
divided into two subsets, to generate a training and a test set. The three normal ROIs obtained from the same mam- mogram as the mass ROI were divided into the same sub-
set as the mass ROI, to ensure the independence of the training and test sets and to allow comparison of the per-
formance of different CAD methods. Standard metrics were computed to reflect the selection and difficulty of cases (26). The rationale for initially using the ROI data-
base, as opposed to the database of full mammograms, was that at the time this work was developed it was easier to
implement a fully automatic method to evaluate the rela- tive performance of each CAD module and its related ROC curves.
475
a. b. c.
d. e. f. Figure 2, Comparison of adaptive and nonadaptive directional wavelet transform modules for directional feature extraction. Images show (a) a raw subimage with a stellate mass, (b) the segmentation result, (c) directional features extracted with the nonadaptive eight- directional directional wavelet transform, (d) directional features extracted with the adaptive directional wavelet transform, which detects more directional features than the transform used for c, (e) spicules identified with nonadaptive directional wavelet transform analysis, and (f) spicules identified with adaptive directional wavelet transform analysis.
Full mammographic image database.--This database, which contained 50 normal and 50 abnormal single-view full mammograms, was used for the purpose of comparing the two overall CAD methods. Both normal and abnormal mammograms manifested typical variations in parenchy- mal density and structure and ranged from very dense to fat breasts. The abnormal mammograms contained at least
one mass of varying size and locations. The database con- tained 20 stellate, 14 circumscribed, and 19 irregular masses, all of which were biopsy-proved cancers. The
original database previously reported for the nonadaptive and Markov random field, or MRF, CAD method con-
tained 16 minimal cancers. Five additional minimal can- cers (difficult cases) were added to this database, as shown in the Table, to allow different comparisons of the relative
performance for the older MRF method and the two CAD methods reported here. For all abnormal mammograms, a reference image or electronic truth file was created, in
which the tumor was labeled by an expert mammographer based on visual criteria and biopsy results. Histograms of the effective size and contrast of the masses in our data- base are shown in Figures 3 and 4, respectively.
RESULTS
Comparison of CAD Modules for Adaptive and Nonadaptive Methods
An example of the improvement in the identification of spiculations with the adaptive directional wavelet trans- form module, compared with the nonadaptive module, is
476
24
22
20
16
C ~, 12 ¢
8
e
4
<11 11- 20 21- 30 31- 40 41- 50 >50 Mass Effective Size (ram)
Histogram of the effective size of the 50 breast Figure 3. masses used in the study. Effective size is defined as the square root of the product of the maximum and minimum dimensions of the mass, as outlined by an expert mammographer.
,o[ 18-
16-
1 <11 11- 20 21- 30 31- 40 41- 50 >50
Mass Contrast (Gray-Scale Value Difference)
Figure 4. Histogram of the contrast enhancement of the 50 breast masses at the 16-bit gray-scale values used in the study. Contrast enhancement is defined here as the difference between the average gray-scale value of the mass and that of the sur- rounding tissues.
Free-Response ROC Curves Comparing the Performance of the MRF, Nonadaptive CAD, and Adaptive CAD Methods
Method
Full Stellate Minimal Database Masses Cancers* (n = 100) (n= 20) (n= 21)
MRF method Detection sensitivity (%) 80 False-positive detection rate 2
Nonadaptive CAD method Detection sensitivity (%) 89 False-positive detection rate 1.91
Adaptive CAD method Detection sensitivity (%) 96 False-positive detection rate 1.71
100 0.26
100 1.42
100 0.81
Note.--Data are based on full mammographic database, *Minimal cancers are defined as lesions smaller than 1 cm.
71 2.1
76 2.07
95 1.12
shown in Figure 2. More spiculations were detected at a higher spatial frequency around the boundary of the
mass. Detailed analysis of the frequency of spiculations poses several logistical problems, as some fine spicula-
tions are hard to observe with a magnifying glass and
screen-film reading on a light box. The comparison of the
adaptive and nonadaptive directional wavelet transform
modules was made by means of ROC analysis. The ROC
curves were generated by varying the output threshold
levels of the fuzzy binary decision tree, as shown in Fig- ure 1. The true-positive detection fraction was defined as
the ratio of the number of correctly classified abnormal
areas to the total number of abnormal ROIs. The corre-
sponding false-positive detection fraction was defined as
the ratio of normal regions misclassified as abnormal ar-
eas to the total number of normal ROIs. The relative per-
formance of the adaptive image-preprocessing modules
can be seen in Figure 5, which shows the computed ROC
area index (Az) values. For example, the adaptive direc-
tional wavelet transform module showed the most signifi-
cant (P = .135) improvement in Az value, which may be
partly attributed to the improvement in removal of the di- rectional features, resulting in better segmentation of sus-
picious areas. The adaptive tree-structured nonlinear filter
477
Figure 5. Computed ROC curves compar- ing the performance of the three image-pre- processing CAD modules, with and without adaptive criteria. The overall effect of the three adaptive modules was statistically sig- nificant. The database was generated by us- ing ROIs drawn to enclose the mass and nor- mal tissue areas (200 ROIs). TSF = tree- structured nonlinear filter, DWT = directional wavelet transform.
0.!
0.8
0.7
0.6 t . .
g~
~" 0.5
0
¢~ 0.4 6 E
0.3
0.2
0.1
o oi,
#
/ s
~ ) e. Nonadaptive method, Az = 0.86
d. With adaptive TSF, Az = 0.87
• With four-channel wavelet transform and network subimage selection, Az = 0.89
b. With adaptive DWT, Az = 0.91
a. With all adaptive modules, Az = 0.95
The P values are computed as follows: P = 0.0145 between a and e. P = 0.262 between c and e. P = 0.135 between b and e. P = 0.418 between d and e,
0.'2 013 014 015 016 017 018 019 Fa l se -Po s i t i v e F r a c t i o n s
module demonstrated a smaller improvement in com- puted Az value. This observation was attributed primarily to an increased reduction of the false-positive detection rate, due to more robust noise removal and the related re-
ductio n of image artifacts potentially generated by the
wavelet transforms. Similarly, the adaptive selection of subimages with the four-channel tree-structured wavelet transform and neural network subimage selection criteria also demonstrated an improvement in Az value. Visual analysis of these subimage data, in turn, showed that the number of subimages selected varied by as many as one
to three within the 16 subimages reconstructed, suggest- ing that adaptive selection of subimages was important. This method may be considered as an alternative ap-
proach to the use of measurements of parenchymal tissue density to optimize CAD methods for a specific subgroup of images. The advantages of the combined use of all adaptive preprocessing modules was well demonstrated
in Figure 5 and stimulated our interest in the evaluation of adaptive CAD methods for full mammographic images.
Comparison of CAD Methods with Full Mammographic Images
The Table shows a comparison of the relative perfor- mance of the previously reported MRF method, the non- adaptive CAD method, and the adaptive CAD method
proposed here. The computed free-response ROC values, which are based on the changes in thresholding values of the fuzzy binary decision tree for the same case database, are shown in Figure 6, in which the data were fitted by a
free-response ROC fit curve program provided by Chak-
raborty (28). The previous MRF method had a reduced sensitivity of detection of 80% (2.0 false-positive results per image). The sensitivity for detection of minimal can- cers was also substantially reduced at 71% (2.1 false-
positive results per image), showing the higher case se- lection involved. The adaptive CAD method demon- strated significantly (P = .0145) better performance than the nonadaptive CAD method. For comparable false-
positive rates, the adaptive CAD method had a sensitivity of 96%; the nonadaptive method, a sensitivity of 89%. The performance of adaptive CAD was also greater for cases of minimal cancer, for which it had a sensitivity of 95%. For the detection of spiculated masses, the adaptive
CAD method was again more sensitive, with 100% sensi- tivity. The low false-positive rate of the adaptive method is noteworthy; one might expect this number to be larger
due to the enhanced sensitivity for the detection of subtle directional features. The low rate may be attributed partly
to the use of adaptive noise suppression to improve im- age artifact suppression prior to the application of these high-order transforms.
478
Vol 6, No 8, August 1999 ADAPTIVE AND NONADAPTIVE CAD METHODS
0.9
0.8
0.7
° ~ -~ 0.6
0.5
0.4
& 0.3
0.2
0.1
0 0
Adaptive method ¢7¢, ¢'~¢'~*~" ...~.-~" " ~" . . . . . . . . .
~ S s ° . - -
s s ,,, , ' ~ , - , -'" Nonadaptive method
j s / " # / * • #a
U ~, . ." ~ MRF method
s • • Z
• Z • , " • J s
• Z" " • # J
t
I i I t i
0.5 1 1.5 2 2.5
False-Positive Resu l t s per I m a g e
Figure 6. Free-response ROC curve of the true-positive mass detection rate versus the false-posit ive detection rate per image for the database shown in the Table. MRF= Markov random field.
The performance of CAD methods for retrospective
case studies of mass detection has been shown to depend on the size of the image database used for testing and the criteria for inclusion of difficult cases (29,30). Similarly,
the performance of CAD methods generally declines
when they are applied to large prospective case studies (31). For this reason, there is a strong consensus that CAD methods, particularly for mass detection, must im- prove subtantially for CAD to be useful for breast cancer
screening or for later application to multicenter trials. For example, Ascher et al (29) recently suggested that a de- tection sensitivity of the order of 90% and a low false- positive rate of 0.1 per image was required. CAD algo-
rithms to date use CAD image-preprocessing modules that have fixed parameters; hence, the performance Of these algorithms is limited because they are not easily op- timized, as discussed at a recent digital mammography
workshop (29). An entirely new class of CAD methods may therefore be necessary to reach these objectives. One
approach, as proposed here, is the use of multiresolution and multiorientation wavelet transforms for improved
feature extraction and the use of adaPtive techniques that
may be less image dependent. In this study, we compared the relative performance of
the CAD modules and overall CAD methods with a retro-
spective database to show the improvement in the CAD algorithms. We have made careful determinations of ground truth by expert mammographers, spatially deter- mined electronically, to allow the automatic generation of the computed free-response ROC curves. The adaptive
CAD method has a number of unique features. First, it is
a modular CAD approach in which each module, such as the image-preprocessing module, can be independently optimized (as shown in Fig 1). Second, the image-prepro-
cessing modules are a first step toward the development of a more generalized CAD method for different digitiz- ers. We have also reported (31) a method for image stan- dardization that--coupled with the adaptive approach--
will help achieve a more general solution for the applica- tion of CAD to different digitizers or direct digital sensors. Further improvements considered for the adaptive meth- od are the use of genetic algorithms for feature selection and other classifiers, such as neural networks with Kal- man filtering for improved convergence (20), instead of
the fuzzy binary decision tree employed here.
IEFERENCE5
1. Giger ML. Computer-aided diagnosis, in: Haus AG, Yaffe M J, eds. Syl- labus: a categorical course in physics--technical aspects of breast im- aging. Oak Brook, II1: Radiological Society of North America, 1993; 283-298.
2. Bassett LW, Bunnell DH, Jahashahi RH, Gold RH, Arndt RD, Linsman L. Breast cancer detection: one versus two views. Radiology 1987; 165:95-97.
479
3. , Sickle EA, Weber WN, Galvin HB. Baseline screening mammography: one vs two views per breast. A JR 1986; 147:1149-1153.
4. Wolverton D, Nishikawa RM, Zouras W, Doi K. CAD in digital mam- mography: computerized detection and classification of masses in: Digital mammography. Amsterdam, the Netherlands: Elsevier Science, 1994.
5. Kegelmeyer WP, Pruneda JM, Bourland PD, Hillis A, Riggs MW, Nip- per ML. Computer-aided mammographic screening for spiculated le- sions. Radiology 1994; 191:331-337.
6. Gupta R, Undrill PE. The use of texture analysis to delineate suspi- cious masses in mammography. Phys Med Bio11995; 40:835-855.
7. Petrick N, Chan HP, Sahiner B, et al. Automated detection of breast masses on digital mammograms using a convolution neural network for morphological and texture classification. Proc World Congress Neural Networks 1995; 2:872-875.
8. Brzakovic D, Luo XN, Brzakovic P. An approach to automated detec- tion of tumors in mar0mograms. IEEE Trans Med Imaging 1990; 9:233- 241.
9. Kegelmeyer WP. Evaluation of stellate lesion detection in a standard mammogram data set. Proc SPIE 1993; 1905:787-793.
10. Huo Z, Giger ML, Vyborny C J, et al. Analysis of spiculation in the com- puterized classification of mammographic masses. Med Phys 1995; 22:1569-1579.
11. Anderberg MR. Clustering analysis for application. New York, NY: Aca- demic Press, 1973.
12. Karssemeijer N. Recognition of stellate lesions in digital mammograms. In: Digital mammography. Amsterdam, the Netherlands: Elsevier Sci- ence, 1994.
13. Karssemeijer N, te Brake GM. Detection of stellate distortion in mam- mograms. IEEE Trans Med Imaging 1996; 15:611-619.
14. Qian W, Li L, Clarke LP. Digital mammography: wavelet based CAD method for mass detection. Proc SPIE Med Imaging 1997; 3034: 790- 801.
15. Li L, Qian W, Clarke LP. Digital mammography: CAD method for mass detection using multiresolution and multiorientation wavelet transforms. Acad Radio11997; 4:724-731.
16. Qian W, Clarke LP, Kallegi M. Digital mammography: M-channel quad- rature mirror filters for microcalcification extraction. Comput Imaging Graphics 1994; 18:301-314.
17. Qain W, Li L, Clarke LP. Image feature extraction for mass detection using digital mammography: influence of wavelet analysis. Med Phys 1999; 26: 402-408.
18. Qian W, Clarke LP, Kallergi M, Clark RA. Tree-structured nonlinear fil- ters in digital mammography. IEEE Trans Med Imaging 1994; 13:25-36.
19. Qian W, Clarke LP, Kallergi M, et al. Application of wavelet transform for image enhancement in medical imaging. In: IntelIigent engineering systems through artificial neural networks. Vol 4. New York, NY: Ameri- can Society of Mechanical Engineers, 1994; 4:651-660.
20. Clarke LP, Zheng B, Qian W. Artificial neural network for pattern recog- nition in mammography. Presented at the World Congress on Neural Networks, Calif, June 4-9, 1994.
21. Li L, Qian W, Clarke LP. X-ray medical image processing using direc- tional wavelet transform. Proc I EEE Int Conf Acoustics Speech Signal Processing 1996; 4.
22. Qian W, Clarke LP, Kallergi M, Zheng B, Clark RA. Wavelet transform for computer assisted diagnosis (CAD) for digital mammography. IEEE Eng Med Bio11995; 14:561-569.
23. Qian W, Clarke LP. Wavelet-based neural network with fuzzy-logic adaptivity for nuclear image restoration. Proc IEEE 1996; 84(special is- sue): 1458-1473.
24. Qian W, Clarke LP. Hybrid M-channel wavelet transform methods: adaptive, automatic and digital X-ray sensor independent. Med Phys 1995; 22:983-984.
25. Pappas TB. An adaptive clustering algorithm for image segmentation. IEEE Trans Signal Process 1992; 40:902-914.
26. Li HD, Kallergi M, Clarke LP, Jain VK, Clark RA. Markov random field for tumor detection in digital mammography. IEEE Trans Med Imaging 1995; 14:565-576.
27. Chan HP, Wei D, Helvie MA, et al. Computer-aided classification of mammographic masses and normal tissue: linear discriminant analysis in texture feature space. Phys Med Bio11995; 40:857-876.
28. Chakraborty DP. Maximum likelihood analysis of free-response re- ceiver operating characteristic (FROC) data. Med Phys 1989; 16:561- 568.
29. Ascher S, et al. Image processing and computer assisted diagnosis (CAD). Federal Multi-Agency Consortium on Imaging Technologies to Improve Women's Health, Technology Transfer Workshop on Breast Cancer Detection, Diagnosis, and Treatment. Office of Women's Health Workshop on Digital Mammography, Washington DC, June 1-2, 1997.
30. Kallergi M, Clark RA, Clarke LP. Medical image databases for CAD applications in digital mammography: design issues. Presented at the Medical Informatics Europe, 14th International Congress, Sithonia, Greece, May 25-29, 1997.
31. Velthuizen RP, Clarke LP. Digitized mammogram standardization for display and CAD. Presented at SPIE, February 21-26, 1998.
4 8 0