Hindawi Publishing CorporationComputational and Mathematical Methods in MedicineVolume 2013 Article ID 593175 8 pageshttpdxdoiorg1011552013593175
Research ArticleSegmentation of the Striatum from MR Brain Images toCalculate the 99mTc-TRODAT-1 Binding Ratio in SPECT Images
Ching-Fen Jiang1 Chiung-Chih Chang2 Shu-Hua Huang3 and Chia-Hsiang Wu1
1 Department of Biomedical Engineering I-Shou University Kaohsiung 82445 Taiwan2Department of Neurology Chang Gung Memorial Hospital Kaohsiung Medical Center Chang Gung University College of MedicineKaohsiung 83301 Taiwan
3Department of Nuclear Medicine Chang Gung Memorial Hospital Kaohsiung Medical CenterChang Gung University College of Medicine Kaohsiung 83301 Taiwan
Correspondence should be addressed to Ching-Fen Jiang cfjiangisuedutw
Received 18 January 2013 Revised 3 June 2013 Accepted 4 June 2013
Academic Editor Norio Tagawa
Copyright copy 2013 Ching-Fen Jiang et alThis is an open access article distributed under theCreativeCommonsAttribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
Quantification of regional 99mTc-TRODAT-1 binding ratio in the striatum regions in SPECT images is essential for differentialdiagnosis between Alzheimerrsquos and Parkinsonrsquos diseases Defining the region of the striatum in the SPECT image is the firststep toward success in the quantification of the TRODAT-1 binding ratio However because SPECT images reveal insufficientinformation regarding the anatomical structure of the brain correct delineation of the striatum directly from the SPECT imageis almost impossible We present a method integrating the active contour model and the hybrid registration technique to extractregions from MR T1-weighted images and map them into the corresponding SPECT images Results from three normal subjectssuggest that the segmentation accuracy using the proposed method was compatible with the expert decision but has a higherefficiency and reproducibility than manual delineation The binding ratio derived by this method correlated well (R2 = 076) withthose values calculated by commercial software suggesting the feasibility of the proposed method
1 Introduction
Alzheimerrsquos and Parkinsonrsquos diseases are two common neu-rodegenerative diseases associated with the aging processThe induced intellectual and functional deterioration ofpatients with these diseases can not only bring a heavyload to hisher family but also has an economic impact onsociety Early diagnosis with appropriate treatment within areasonable time frame can prevent abrupt degeneration ofthese diseases and distressing symptoms
The current trend in the early diagnosis of such diseasesis usually to adopt a combination of functional imagesand structural images to inspect the functional and struc-tural changes in specific brain regions However qualitativeobservation alone limits early detection of neurodegener-ative diseases because the associated functionalstructuralchanges are slowly progressive in the early stage and canbe too subtle to be detected by human vision Therefore
quantification of these changes can facilitate early detectionof neurodegenerative diseases
SPECT imaging of dopamine transporter with 99mTc-TRODAT-1 (TRODAT-1) has been proposed to be a valu-able and feasible means for the diagnosis of Parkinsonrsquosdisease and dementia with Lewy bodies (DLB) [1ndash5] Thespecific tracer TRODAT-1 a radiolabeled tropane that bindsdopamine transporters allows in vivo assessment of thepresynaptic dopaminergic neuron activity inside the striatum[3 6] SPECT images from patients with these diseases reveala decrease in specific striatal uptake of TRODAT-1 in termsof a dull contrast of radioactivity between the striatum andadjacent brain tissue due to a selective loss of dopamine in thestriatum Even though several approaches show the feasibilityof using TRODAT-1 SPECT in the evaluation of patientsin the early stages of these neurodegenerative diseasesvisual inspection or semi-auto quantification cannot avoidhigh intra- or interobserver variability and thus hampers
2 Computational and Mathematical Methods in Medicine
the associated diagnostic accuracy [7 8] A reliable automaticmethod could considerably speed up the procedure andmakeit more reproducible
Even though some commercialized software packagesprovide automatic calculation of theTRODAT-1 binding ratio(BR) definition of the striatum in the SPECT image stillrelies on manual delineation However the brain structure ispoorly-defined in SPECT images which reveal more func-tional information than anatomical structural informationTherefore demarcation of the region of interest (ROI) inthe SPECT image is usually carried out by overlapping theSPECT images with the corresponding MR images such thatphysicians can map the ROI delineated in the MR images tothe SPECT images Within this process there are two keycomponents of determining the accuracy of the TRODAT-1binding quantification First the striatum should be correctlydefined Second the MR images must be precisely registeredwith the corresponding SPECT images However even awell-trained physician can hardly guarantee obtaining accurateand repeatable results at these two stages Therefore thisstudy aims to develop a robust method to fulfill the shortagesin the current approaches
Regarding the segmentation task for subcortical brainstructures several semiautomatic methods have been pro-posed Worth et al proposed the regional thresholdingmethod to segment the caudate from the adjacent tissue [9]A box was manually located to cover these three tissuesincluding the ventricles the caudate and some white matterto derive a bimodal histogram and then the threshold wasdetermined as the mean of the two peaks of the histogramHowever the box location required human determinationand the vague boundaries of the caudate tail surroundedby gray matter still require manual drawing Barra andBoire proposed a fuzzy-logic-based method to segmentsubcortical brain structures in MR images by integratingthe numerical information derived from the wavelet featuresand structural information containing symbolized distanceand relative direction coding [10] More recently Xia et altook advantage of the high-contrast lateral ventricle as thereference to localize the upper and lower bonds of the caudatenucleus for region growing [11] Fine-tuning according tothe topological and morphological information was stillrequired to smooth the initial segmentation In view of thesemethods as several factors such as the complex anatomicbrain structure the connection of different tissues of a similarintensity the heterogeneous intensity within the same class oftissue and the partial volume effect limit the performanceof fully automatic segmentation of the striatum thereforeusing expert knowledge to refine the initial ROI derivedby running the computer program was inevitable Howevervisual confirmation and manual correction are conductedslice by slice and thus may still be labor intensive and timeconsuming
Instead of applying expert knowledge in the last step torefine the segmentation in the previous studies we proposea new approach using an active contour model to reverse theprocess of segmentation that is to let the expert determinethe rough location of the striatum and allow the computer toperform the refinement such that human intervention can be
minimized and the segmentation efficiency can be enhancedThe segmented regions were then mapped into the corre-sponding SPECT images via a hybrid registrationmethod forBR calculation These methods associated with the imagingprotocol are described in detail in Section 2 To verify thereliability of the proposed method the segmentation resultsand the derivative BRs were compared with those of expertsassisted by commercial software The results are presentedand discussed following which a brief conclusion is made
2 Methods
We used hybrid SPECTCT and 3D T1-weighted images toachieve the goal Each volume played a distinct role in theoverall process The registration of the MR and the SPECTvolume pairs was first conducted using the corresponding CTvolume as a medium After that the striatum was segmentedfrom the registered MR images Once the MR images wereadjusted to the same size under the same coordinates withthe SPECT images through registration the ROIs obtainedby applying the active contour model to the registered MRimages could be directly mapped into the SPECT imagesto calculate the binding potential The overall process issummarized in Figure 1 and described in detail below
21 Imaging Protocol For this examination all the patientswere injected intravenously with a single bolus dose of740MBq (20mCi) of 99mTc-TRODAT-1 Brain SPECTCT(Symbia T Siemens Erlangen Germany) images wereobtained 4 hours laterThe SPECTCT scanner was equippedwith low-energy high-resolution collimators and a dual-slicespiral CT Acquisition parameters for SPECT were a 128 times128 matrix 500mm FOV with 60 frames (40 sframe) Thescan parameters for the CT were 130 kV 17mAs 5mm slicesand image reconstruction with a medium-smooth kernelThe SPECT images were attenuation-corrected based on theCT images and scatter-corrected with Flash 3DR algorithm(ordered subsets expectation and 3D maximization withresolution correction) with 8 subsets and 8 iterations
MR images were acquired using a 30 T MRI scanner(Excite GE Medical Systems Milwaukee WI USA) Struc-tural images were acquired for an anatomical referenceusing a T1-weighted inversion-recovery-prepared three-dimensional spoiled and gradient-recalled acquisition in asteady-state sequence with repetition timeinversion time =8600ms450ms a 240 times 240mm field of view and a 1mmslice thickness
22 Image Registration To precisely map the ROI delineatedfrom the MR into a corresponding position in the SPECTimage registration of the MR volume with the SPECT-CT volume was required Even though several automaticregistration methods have been proposed their success isonly guaranteed when the two scanning data to be registeredcontain consistent volumes However the clinical volumesets from different image modalities are usually truncatedunevenly lending additional difficulties to the applicationof conventional registration methods such as principal axes
Computational and Mathematical Methods in Medicine 3
Vertical axes alignment
Voxel size adjustment
Derivation of the principle axes
Generalized Hough transform
ROI segmentation
Fine-tune
Coarse registration
Original images
Detect MCSA
Search for theoptimal parameters
MR imagesHead segmentationCT images
SPECT images
SPECT CT scan
The active contour model gradient-vector flow
Rotate and shift
Map MR ST regions into SPECTto calculate binding ratio
Figure 1 The overall process to derive the BRs in SPECT images via registration of the images from SPECT-CT and MR with ROIsegmentation from the registered MR images
registration (PAR) or mutual information (MI) To alleviatethis problem we developed a hybrid registration methodcombining principal axes registrationwith the generalHoughtransform [12] In addition we took advantage of SPECT-CTwhich can acquire SPECT andCT images simultaneouslywhile the patient maintains hisher position on the samecouch Registration used the CT image as the registrationmedium to increase the registration accuracy between theSPECT and the MR image volumes The registration processis fully automatic The essential idea of the design is brieflydescribed below
The voxel size was adjusted to a 1mm3 cube throughbicubic interpolation prior to the following registrationprocess The 3D head was segmented as an entity to deriveits three principal axes prior to registration In this two-stage registration scheme principal axes registration was firstapplied for coarse registration followed by fine-tuning viaapplying the general Hough transform to the contour of themaximal cross-sectional area (MCSA) The original conceptof principal axes registration (PAR) is to superimpose thetwo volumes by aligning the corresponding three principalaxes from both head volumes [13] However the registrationaccuracy of PAR is restricted by the degree of correspondencebetween the two sets of principal axes [14] As the scanningrange of one image modality is usually not the same as theother the centroids of the two different volume sets wouldnot be identical In consequence the two sets of principalaxes derived from the different centroids do not coincidewith each other Therefore in the coarse-registration stage
we only adopted PAR to adjust the orientations of the longaxis of the head to be parallel with the 119911-axis of the systemcoordinates After this stage the long axis from both headvolumes coincided with each other but the horizontal planeswith the two short axes from the two volumes were stillmismatched
In the second stage the registration error in the hori-zontal plane was then fine-tuned We then turned the 3Dregistration task into a 2D one by searching for the slicescontaining the (MCSA) in both volumes in that we hadproved the reliability of using the MCSA as the anatomicalfeature for registration [15] The vertical shift was firstcorrected by aligning these two slices then the detectedcontour of the MCSA was used to derive the registrationparameters via the generalized Hough transform (GHT)Theprocess of the GHT algorithm in this approach includedtwo steps First an R-table was built by calculating thevector set 997888119886
119894 between each contour point (119909
119894 119910119894) and
the center of the contour 119875119888(119909119888 119910119888) in the CT image
Then the corresponding center point (1198751015840119888) was derived by
searching for the maximal intersection via remapping thevector information to each contour point X
119894(119909119894 119910119894) in the
MR image In this study as there was no scale for referenceand the voxel size had been adjusted to be the same weadapted a robust search only for the rotation angle 120573 in (1)when the optimal match between 119875
119888and 1198751015840
119888was achieved
119909119888= 119909119894+ 120574 cos (120579 + 120573)
119910119888= 119910119894+ 120574 sin (120579 + 120573)
(1)
4 Computational and Mathematical Methods in Medicine
Caudate nucleus
Putamen
Left striatum of basal ganglia
Lateral ventricle
Figure 2 Anatomical structure of the striatum from the axial viewof an MR T1-weighted image
where 120579 is the angle between the directional vector997888119886119894and the
positive direction of the 119909-axis and 120574 is the length of the 997888119886119894
The registration parameters of the rigid transformderivedabovewere then applied to theMR volumes tomatchwith theSPECT images
23 ROI Segmentation from MR T1-Weighted Images Theregistered MR T1-weighted images obtained from the pre-vious stage were then used as the reference to demarcatethe striatum on the corresponding SPECT images As theassessment of TRODAT-1 BR is usually carried out from theaxial view of SPECT images the segmentation of the striatumwas performed in the axial planes of the MR images Figure 2shows the structure of the striatum from the axial view of theMR image It can be seen that the left and right sides of stria-tum of the basal ganglia are located beside the ventricle Eachside of the striatum can be further divided into the caudatenucleus and putamen We named the two pairs of caudatenucleus and putamen the ST regions However the divisionof the ST regions is not obvious because they usually fuse withother brain structures The unclear cut between the caudatenucleus and putamen and the surrounding brain structurebrings up difficulties in isolating the ST regions solely usingautomatic image segmentation techniqueswithout any expertintervention
To segment these four ROIs we adopted a modifiedactive contour model In this way an initial contour of thefirst slice can be determined by an expert according to thetopological andmorphological characteristics of the STOncethe location and shape of the ST regions are confined into thebond of the initial contour then refinement can be carriedout by the computer according to the intensity informationIn addition assuming smooth variation of the 3D ST regioncontour the final contour of the present slice can be directlyused as the initial contour for the next slice To achieve thisgoal the active contour model is a suitable choice
The basis of the active contour model named snake isto represent an initial contour in the parametric form of
V(119904) = [119909(119904) 119910(119904)] 119904 isin [0 1] that deforms to the optimalshape by minimizing the energy functional
119864snake = int1
0
[119864int (V (119904)) + 119864ext (V (119904))] 119889119904
where 120572 and 120573 are the parameters to weight the influence onthe curve deformation from the curversquos tension V1015840(119904) and therigidity V10158401015840(119904) respectively
Theoretically at the minima of the energy functional thesnake must satisfy the Euler equation
120572V10158401015840(119904) minus 120573V
1015840101584010158401015840(119904) minus nabla119864 ext(V (119904)) = 0 (3)
As the first derivative of energy gives the force the aboveequation can be interpreted as a force balance equation
119865int (V) + 119865ext (V) = 0 (4)
The internal force 119865int(V) = 120572V10158401015840(119904)minus120573V1015840101584010158401015840(119904) restricts the
curve to stretch and bend while the external force 119865ext(V) =minusnabla119864 ext(V) pulls the curve toward the desired image edges
The snake is an active rather than a salient model dueto the dynamic deformation process by treating the forcebalance equation as function of time 119905Therefore the solutionof (3) can be approximated by iteratively searching for thesteady state of the following equation where the V(119904 119905) =[119909(119904 119905) 119910(119904 119905)] denotes V(119904) at the 119905th iteration
120597V (119904 119905)
120597119905
= 120572V10158401015840(119904 119905) minus 120573V
1015840101584010158401015840(119904 119905) minus nabla119864 ext(V (119904 119905)) (5)
In practice a numerical solution to (5) can be achieved bydiscretizing 119904 iteratively using a finite difference method [16]as per
x119905= (A + 120574I)minus1 (x
119905minus1minus p119909119905minus1)
y119905= (A + 120574I)minus1 (y
119905minus1minus p119910119905minus1)
(6)
where A is a pentadiagonal matrix containing the constants120572 and 120573 The parameter of 120574 is the step size to control thedegree of the contour deformation between iterations I isthe unit matrix x
119905and y119905are the vectors consisting of the 119909-
and 119910-coordinates of the contour V(119904 119905) respectively p119909119905minus1
and p119910119905minus1
are the vectors containing 120597119864ext(119909(119904 119905 minus 1) 119910(119904 119905 minus1))120597119909 and 120597119864ext(119909(119904 119905 minus 1) 119910(119904 119905 minus 1))120597119910 as their elementsfor all 119904 respectively
The external force (nabla119864ext) in the active model can usuallybe classified into two types static and dynamic Static forcesare derived from the image gradients which do not changethroughout the deformation process while dynamic forcesvary as the snake deforms Using the image gradient as theexternal force makes the conventional snake difficult to moveinto a concave edge because the null image gradients withina homogenous region inside the contour fail to attract thecontour and as a result the contour is only affected by
Computational and Mathematical Methods in Medicine 5
the internal forces Even though several dynamic externalforces have been proposed to alleviate such a limitation ofthe static external forces they also raised other problemsincreasing the calculation complexity or leading to uncon-trollable deformation [17 18] A new static external forcecalled gradient-vector flow (GVF) adding the directionalproperty into the original image gradient map was proposedby Xu and Prince to improve the performance of the staticsnake in concave edge detection [19] Several reports havedemonstrated the success of applying the GVF snake tomedical image segmentation [20ndash23] including brain MRI[24] This encouraged us to apply the GVF to segment the STregions in our study
The gradient-vector-flow field is defined as k(119909 119910) =[119906(119909 119910) V(119909 119910)] such that the external energy functionbecomes
where 120583 is a parameter to control the degree of smoothness ofthe gradient-vector-flow field and nabla119891 is an edge map derivedfrom the original image 119891(119909 119910)
To solve the equation numerically by discretization anditeration let 119899 indicate the times of iteration and theincrements in 119909 119910 and 119905 are all equal to 1 The relation ofvector flows from the current to the next position can bederived as
There are 4 parameters determined empirically to obtainthe optimal results in the approach The elasticity parameter(120572) and the rigidity parameter (120573) in (2) were set to be 01 and02 respectively The parameter (120574) in (6) was set to be 1 Theexternal force weight (120583) in (7) was set to be 05
3 Results and Discussion
31 Registration Results An example is given in Figure 3 toillustrate the use of our developed interface to detect the slicescontaining the MCSA from the CT and MR volumes andregister these two images through the GHT Once the rigidtransform with the registration parameters had been appliedto the MR image volume it can directly match the SPECTvolume as shown in Figure 4 The registration accuracyreached 9648 Our previous study quantitatively evaluated
Figure 3 An example of registration of the CT image (middle smallpanel) and theMR image (right small panel) to render the final fusedimage (left large panel)The red and blue lines in the small panels arethe detected boundaries and the contour of the head
Table 1 Correspondence of the manual delineation between theobservers
JI () Rater A-Rater B Rater B-Rater C Rater A-Rater CCase 1 604 plusmn 48 621 plusmn 11 756 plusmn 21
Case 2 543 plusmn 45 547 plusmn 56 718 plusmn 52
Case 3 580 plusmn 47 581 plusmn 54 766 plusmn 52
the registration accuracy of the proposed method better thanthe results obtained solely using the PAR method or directlyregistering SPECT with MR images [12]
32 Segmentation Results The expert delineation and theGVF segmentation of the ST regions containing two pairsof the caudate nucleus and putamen are given in Figure 5 Aquantitative comparison of these twomethods is given below
We used the Jaccard index (JI) to quantify the degree ofmatch between the two corresponding ROIsThe JI is definedas the ratio of the intersection of two volumes Ω
1and Ω
2by
the union of them If the two volumes completely overlap theJI value is equal to 100
JI =1003816100381610038161003816Ω1cap Ω2
1003816100381610038161003816
1003816100381610038161003816Ω1cup Ω2
1003816100381610038161003816
times 100 (9)
The five sequential axial slices containing the ST fromthree normal cases were recruited in the comparative eval-uation Three neurologists first manually delineated the STregions including the caudate nucleus and putamen on twolateral sides of the brain The intrarater correspondences interms of the mean and standard deviation of the JI valuesfrom the five slices are listed in Table 1 suggesting greatdifferences between observers It was found that Raters A andChad the highest correspondence with a JI value greater than70
The JI values were also derived by mapping the manuallydefined contours by each rater into the GVF segmentedresults The initial contour of the first slice in each case was
6 Computational and Mathematical Methods in Medicine
(a) (b) (c)
Figure 4 Registration between (a) SPECT and (b) MR to give the final overlaid image in (c)
(a) (b) (c)
Figure 5 The ST regions in (a) the original MR T1-weighted image and the corresponding segmentation results by (b) manual delineationand (c) the GVF snake
Table 2 Correspondence betweenmanual delineation and theGVFsnake result for each observer
JI () Rater A-GVFsnake
Rater B-GVFsnake
Rater C-GVFsnake
Case 1 644 plusmn 90 562 plusmn 55 653 plusmn 48
Case 2 686 plusmn 16 577 plusmn 51 654 plusmn 65
Case 3 611 plusmn 38 515 plusmn 31 599 plusmn 37
defined by the same specialist in the GVF snake processTable 2 shows the correspondence with GVF segmentationWe used the paired 119905-test to evaluate the significance levelbetween the JIs derived from the interrater comparison andthose from the rater-GVF comparison for each slice in eachcase The insignificant differences (119875 = 0124 under a 95confidence interval) suggest that the segmentation accuracyusing the GVF snake is compatible with the level of manualdrawing
Rater A the chief neurologist was required to conductthe manual drawing twice The JI values of the two delin-eations are listed in the middle column of Table 3 showing
Table 3 Correspondence between two repeated conductions ofeach method
that the correspondence declined along with the slice num-ber Instead of segmentation solely by hand the GVF snakewas also applied twice to the same set of images Only thefirst slice required an initial contour manually defined by therater each time The JI values of repeated conduction of theGVF snake are also listed in the third column of Table 3suggesting more stable results than those from slice-by-slicemanual drawing
Computational and Mathematical Methods in Medicine 7
In comparison with the index of overlap (similar to JI)between hand-drawing and computer-aided segmentationreported in the literature [9ndash11] the JI values obtained inour study were relatively low This could be due to the extraregion that is the putamen involved in our study Thesegmentation target of the previous reports is focused onthe caudate nucleuses that are next to ventricles with greatercontrast (Figure 1) and therefore more easily identified Incomparison with the caudate nucleuses the low contrastof the putamen to the surrounding tissue increases thedifficulty of extraction Using expert hand drawing as thecomparison basis seems to be the only choice in currentstudies since there is no gold standard to determine theabsolute accuracy of segmentation of the ST regions dueto individual-dependent variation in the brain structureHowever we demonstrated that significant interobserver andintraobserver variability in such a decision exists even amongthe well-trained neurologists participating in our studywhich was overlooked in previous studies The inconsistencyin decision-making could be incurred by the small size (inthe order of 100 pixels) of the structures as compared with theimaging resolution and image noise Under a compatible levelof precision as shown in Tables 1 and 2 we demonstrated thatthe reproducibility and consistency improved when using theGVF snake segmentationmethod In addition to stability theGVF snake can save labor and provide a more efficient waythan previous studies to define the ST regions contours inconsecutive slices as it only requires an initial contour drawnby hand in the first slice
33 Binding Ratio Calculation In the final stage of evaluationof the reliability of the proposedmethod after completing theMR and SPECT image registration the BRs were also derivedfrom the segmented ST regions in the SPECT images usingthe proposed method to compare with those obtained usingcommercial software (Siemens Medical Systems KnoxvilleTN USA) in which the ST regions were manually outlinedby an expert The BR was calculated by normalizing themean intensity in the ST regions by the mean intensity inthe occipital cortices Linear regression analysis (Figure 6)revealed a close correlation (CC = 0874 under 95confidence interval) between the BRs derived by the twosystems
4 Conclusions
To calculate the regional TRODAT-1 binding ratio in SPECTstudies accurate and repeatable extraction of the ST regionsfrom MR images is required to indirectly define the cor-responding regions in the SPECT images Segmentationdirectly on the SPECT image is not applicable in this casebecause it distorts the ST regions Clinical routine tends toapply manual delineation of the ST regions which is proneto errors incurred through interobserver and intraobservervariability Previous researchers have developed several seg-mentation algorithms to complete similar tasks where expertdecisions for anatomical and morphological informationwere still necessary to refine the results As the localization
1
11
12
13
14
15
16
1 12 14 16 18 2 22 24 26
Our
met
hod
Commercial software
y = 0466x + 0415
R2= 0764
Correlation coefficient = 0874
Figure 6 Linear regression analysis between the BRs derived by ourmethod and those by the commercial software
of the ST regions is a knowledge-driven task the proposedmethod allowed the expert to assign the initial contourin the proper location and applied the gradient-vector-flow snake to approach the real contours In such a waythe complexity of the algorithm can be reduced and theefficiency of segmentation can be increased Results fromthree normal subjects showed a higher reproducibility ofthe proposed method than manual segmentation undercompatible segmentation accuracy The MR images withsegmented ST regions were overlaid on the SPECT imagesusing our previously developed registration algorithm tocalculate the TRODAT-1 BRThe derived BRs correlated wellwith those derived using commercial software suggesting agood reliability of the proposed method
Acknowledgments
The authors would like to thank Dr Chen Nai-Ching andDr Chi-Wei Huang from the Department of NeurologyChang Gung Memorial Hospital Kaohsiung Medical CenterKaohsiung Taiwan for their kind assistance in the task ofmanual drawing of the ST regions in this study
References
[1] S Asenbaum ldquoNuclear medicine in neurodegenerative disor-dersrdquo Imaging Decisions MRI vol 6 pp 19ndash28 2002
[2] K L Chou H I Hurtig M B Stern et al ldquoDiagnostic accuracyof [ 99mTc]TRODAT-1 SPECT imaging in early Parkinsonrsquosdiseaserdquo Parkinsonism and Related Disorders vol 10 no 6 pp375ndash379 2004
[3] J L Cummings C Henchcliffe S Schaier T Simuni AWaxman and P Kemp ldquoThe role of dopaminergic imaging inpatients with symptoms of dopaminergic system neurodegen-erationrdquo Brain vol 134 no 11 pp 3146ndash3166 2011
[4] A Siderowf A Newberg K L Chou et al ldquo[ 99mTc]TRODAT-1 SPECT imaging correlates with odor identification in earlyParkinson diseaserdquo Neurology vol 64 no 10 pp 1716ndash17202005
8 Computational and Mathematical Methods in Medicine
[5] Y-H Weng T-C Yen M-C Chen et al ldquoSensitivity andspecificity of 99mTc-TRODAT-1 SPECT imaging in differenti-ating patients with idiopathic Parkinsonrsquos disease from healthysubjectsrdquo Journal of NuclearMedicine vol 45 no 3 pp 393ndash4012004
[6] W-S Huang S-Z Lin J-C Lin S-P Wey G Ting and R-SLiu ldquoEvaluation of early-stage Parkinsonrsquos disease with 99mTc-TRODAT-1 imagingrdquo Journal of NuclearMedicine vol 42 no 9pp 1303ndash1308 2001
[7] P D Acton P T Meyer P D Mozley K Plossl and H F KungldquoSimplified quantification of dopamine transporters in humansusing [ 99mTc]TRODAT-1 and single-photon emission tomog-raphyrdquo European Journal of Nuclear Medicine and MolecularImaging vol 27 no 11 pp 1714ndash1718 2000
[8] P D Acton P D Mozley and H F Kung ldquoLogistic discrimi-nant parametric mapping a novel method for the pixel-baseddifferential diagnosis of Parkinsonrsquos diseaserdquo European Journalof Nuclear Medicine and Molecular Imaging vol 26 no 11 pp1413ndash1423 1999
[9] A J Worth N Makris M R Patti et al ldquoPrecise segmentationof the lateral ventricles and caudate nucleus in mr brain imagesusing anatomically driven histogramsrdquo IEEE Transactions onMedical Imaging vol 17 no 2 pp 303ndash310 1998
[10] V Barra and J-Y Boire ldquoAutomatic segmentation of subcorticalbrain structures in MR images using information fusionrdquo IEEETransactions on Medical Imaging vol 20 no 7 pp 549ndash5582001
[11] Y Xia K Bettinger L Shen and A L Reiss ldquoAutomaticsegmentation of the caudate nucleus from human brain MRimagesrdquo IEEE Transactions on Medical Imaging vol 26 no 4pp 509ndash517 2007
[12] C-F Jiang C-C Chang and S-H Huang ldquoRegions of interestextraction from SPECT images for neural degeneration assess-ment using multimodality image fusionrdquo MultidimensionalSystems and Signal Processing vol 23 pp 437ndash449 2012
[13] A P Dhawan ldquoImage registrationrdquo in Medical Image AnalysisM Akay Ed pp 251ndash276 IEEE Press Piscataway NJ USA2003
[14] A P Dhawan L K Arata A V Levy and J Mantil ldquoIterativeprincipal axes registrationmethod for analysis ofMR-PETbrainimagesrdquo IEEE Transactions on Biomedical Engineering vol 42no 11 pp 1079ndash1087 1995
[15] C-F Jiang C-H Huang and S-T Yang ldquoUsingmaximal cross-section detection for the registration of 3D image data of theheadrdquo Journal of Medical and Biological Engineering vol 31 no3 pp 217ndash226 2011
[16] MKass AWitkin andD Terzopoulos ldquoSnakes active contourmodelsrdquo International Journal of Computer Vision vol 1 no 4pp 321ndash331 1988
[17] L D Cohen and I Cohen ldquoFinite-element methods for activecontour models and balloons for 2-D and 3-D imagesrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol15 no 11 pp 1131ndash1147 1993
[18] B Leroy I Herlin and L Cohen ldquoMulti-resolution algorithmsfor active contour modelsrdquo in Proceedings of the 12th Inter-national Conference on Analysis and Optimization of SystemsImages (ICAOS rsquo96) pp 58ndash65 1996
[19] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998
[20] B Erkol R H Moss R J Stanley W V Stoecker and EHvatum ldquoAutomatic lesion boundary detection in dermoscopy
images using gradient vector flow snakesrdquo Skin Research andTechnology vol 11 no 1 pp 17ndash26 2005
[21] J Tang and S T Acton ldquoVessel boundary tracking for intravitalmicroscopy via multiscale gradient vector flow snakesrdquo IEEETransactions on Biomedical Engineering vol 51 no 2 pp 316ndash324 2004
[22] J Tang S Millington S T Acton J Crandall and S HurwitzldquoSurface extraction and thickness measurement of the articularcartilage fromMR images using directional gradient vector flowsnakesrdquo IEEE Transactions on Biomedical Engineering vol 53no 5 pp 896ndash907 2006
[23] N Tanki K Murase M Kumashiro et al ldquoQuantification ofleft ventricular volumes from cardiac cine MRI using activecontour model combined with gradient vector flowrdquo MagneticResonance in Medical Sciences vol 4 no 4 pp 191ndash196 2005
[24] C Xu and J L Prince ldquoGradient vector flow deformable mod-elsrdquo in Handbook of Medical Imaging pp 159ndash169 AcademicPress 2000
2 Computational and Mathematical Methods in Medicine
the associated diagnostic accuracy [7 8] A reliable automaticmethod could considerably speed up the procedure andmakeit more reproducible
Even though some commercialized software packagesprovide automatic calculation of theTRODAT-1 binding ratio(BR) definition of the striatum in the SPECT image stillrelies on manual delineation However the brain structure ispoorly-defined in SPECT images which reveal more func-tional information than anatomical structural informationTherefore demarcation of the region of interest (ROI) inthe SPECT image is usually carried out by overlapping theSPECT images with the corresponding MR images such thatphysicians can map the ROI delineated in the MR images tothe SPECT images Within this process there are two keycomponents of determining the accuracy of the TRODAT-1binding quantification First the striatum should be correctlydefined Second the MR images must be precisely registeredwith the corresponding SPECT images However even awell-trained physician can hardly guarantee obtaining accurateand repeatable results at these two stages Therefore thisstudy aims to develop a robust method to fulfill the shortagesin the current approaches
Regarding the segmentation task for subcortical brainstructures several semiautomatic methods have been pro-posed Worth et al proposed the regional thresholdingmethod to segment the caudate from the adjacent tissue [9]A box was manually located to cover these three tissuesincluding the ventricles the caudate and some white matterto derive a bimodal histogram and then the threshold wasdetermined as the mean of the two peaks of the histogramHowever the box location required human determinationand the vague boundaries of the caudate tail surroundedby gray matter still require manual drawing Barra andBoire proposed a fuzzy-logic-based method to segmentsubcortical brain structures in MR images by integratingthe numerical information derived from the wavelet featuresand structural information containing symbolized distanceand relative direction coding [10] More recently Xia et altook advantage of the high-contrast lateral ventricle as thereference to localize the upper and lower bonds of the caudatenucleus for region growing [11] Fine-tuning according tothe topological and morphological information was stillrequired to smooth the initial segmentation In view of thesemethods as several factors such as the complex anatomicbrain structure the connection of different tissues of a similarintensity the heterogeneous intensity within the same class oftissue and the partial volume effect limit the performanceof fully automatic segmentation of the striatum thereforeusing expert knowledge to refine the initial ROI derivedby running the computer program was inevitable Howevervisual confirmation and manual correction are conductedslice by slice and thus may still be labor intensive and timeconsuming
Instead of applying expert knowledge in the last step torefine the segmentation in the previous studies we proposea new approach using an active contour model to reverse theprocess of segmentation that is to let the expert determinethe rough location of the striatum and allow the computer toperform the refinement such that human intervention can be
minimized and the segmentation efficiency can be enhancedThe segmented regions were then mapped into the corre-sponding SPECT images via a hybrid registrationmethod forBR calculation These methods associated with the imagingprotocol are described in detail in Section 2 To verify thereliability of the proposed method the segmentation resultsand the derivative BRs were compared with those of expertsassisted by commercial software The results are presentedand discussed following which a brief conclusion is made
2 Methods
We used hybrid SPECTCT and 3D T1-weighted images toachieve the goal Each volume played a distinct role in theoverall process The registration of the MR and the SPECTvolume pairs was first conducted using the corresponding CTvolume as a medium After that the striatum was segmentedfrom the registered MR images Once the MR images wereadjusted to the same size under the same coordinates withthe SPECT images through registration the ROIs obtainedby applying the active contour model to the registered MRimages could be directly mapped into the SPECT imagesto calculate the binding potential The overall process issummarized in Figure 1 and described in detail below
21 Imaging Protocol For this examination all the patientswere injected intravenously with a single bolus dose of740MBq (20mCi) of 99mTc-TRODAT-1 Brain SPECTCT(Symbia T Siemens Erlangen Germany) images wereobtained 4 hours laterThe SPECTCT scanner was equippedwith low-energy high-resolution collimators and a dual-slicespiral CT Acquisition parameters for SPECT were a 128 times128 matrix 500mm FOV with 60 frames (40 sframe) Thescan parameters for the CT were 130 kV 17mAs 5mm slicesand image reconstruction with a medium-smooth kernelThe SPECT images were attenuation-corrected based on theCT images and scatter-corrected with Flash 3DR algorithm(ordered subsets expectation and 3D maximization withresolution correction) with 8 subsets and 8 iterations
MR images were acquired using a 30 T MRI scanner(Excite GE Medical Systems Milwaukee WI USA) Struc-tural images were acquired for an anatomical referenceusing a T1-weighted inversion-recovery-prepared three-dimensional spoiled and gradient-recalled acquisition in asteady-state sequence with repetition timeinversion time =8600ms450ms a 240 times 240mm field of view and a 1mmslice thickness
22 Image Registration To precisely map the ROI delineatedfrom the MR into a corresponding position in the SPECTimage registration of the MR volume with the SPECT-CT volume was required Even though several automaticregistration methods have been proposed their success isonly guaranteed when the two scanning data to be registeredcontain consistent volumes However the clinical volumesets from different image modalities are usually truncatedunevenly lending additional difficulties to the applicationof conventional registration methods such as principal axes
Computational and Mathematical Methods in Medicine 3
Vertical axes alignment
Voxel size adjustment
Derivation of the principle axes
Generalized Hough transform
ROI segmentation
Fine-tune
Coarse registration
Original images
Detect MCSA
Search for theoptimal parameters
MR imagesHead segmentationCT images
SPECT images
SPECT CT scan
The active contour model gradient-vector flow
Rotate and shift
Map MR ST regions into SPECTto calculate binding ratio
Figure 1 The overall process to derive the BRs in SPECT images via registration of the images from SPECT-CT and MR with ROIsegmentation from the registered MR images
registration (PAR) or mutual information (MI) To alleviatethis problem we developed a hybrid registration methodcombining principal axes registrationwith the generalHoughtransform [12] In addition we took advantage of SPECT-CTwhich can acquire SPECT andCT images simultaneouslywhile the patient maintains hisher position on the samecouch Registration used the CT image as the registrationmedium to increase the registration accuracy between theSPECT and the MR image volumes The registration processis fully automatic The essential idea of the design is brieflydescribed below
The voxel size was adjusted to a 1mm3 cube throughbicubic interpolation prior to the following registrationprocess The 3D head was segmented as an entity to deriveits three principal axes prior to registration In this two-stage registration scheme principal axes registration was firstapplied for coarse registration followed by fine-tuning viaapplying the general Hough transform to the contour of themaximal cross-sectional area (MCSA) The original conceptof principal axes registration (PAR) is to superimpose thetwo volumes by aligning the corresponding three principalaxes from both head volumes [13] However the registrationaccuracy of PAR is restricted by the degree of correspondencebetween the two sets of principal axes [14] As the scanningrange of one image modality is usually not the same as theother the centroids of the two different volume sets wouldnot be identical In consequence the two sets of principalaxes derived from the different centroids do not coincidewith each other Therefore in the coarse-registration stage
we only adopted PAR to adjust the orientations of the longaxis of the head to be parallel with the 119911-axis of the systemcoordinates After this stage the long axis from both headvolumes coincided with each other but the horizontal planeswith the two short axes from the two volumes were stillmismatched
In the second stage the registration error in the hori-zontal plane was then fine-tuned We then turned the 3Dregistration task into a 2D one by searching for the slicescontaining the (MCSA) in both volumes in that we hadproved the reliability of using the MCSA as the anatomicalfeature for registration [15] The vertical shift was firstcorrected by aligning these two slices then the detectedcontour of the MCSA was used to derive the registrationparameters via the generalized Hough transform (GHT)Theprocess of the GHT algorithm in this approach includedtwo steps First an R-table was built by calculating thevector set 997888119886
119894 between each contour point (119909
119894 119910119894) and
the center of the contour 119875119888(119909119888 119910119888) in the CT image
Then the corresponding center point (1198751015840119888) was derived by
searching for the maximal intersection via remapping thevector information to each contour point X
119894(119909119894 119910119894) in the
MR image In this study as there was no scale for referenceand the voxel size had been adjusted to be the same weadapted a robust search only for the rotation angle 120573 in (1)when the optimal match between 119875
119888and 1198751015840
119888was achieved
119909119888= 119909119894+ 120574 cos (120579 + 120573)
119910119888= 119910119894+ 120574 sin (120579 + 120573)
(1)
4 Computational and Mathematical Methods in Medicine
Caudate nucleus
Putamen
Left striatum of basal ganglia
Lateral ventricle
Figure 2 Anatomical structure of the striatum from the axial viewof an MR T1-weighted image
where 120579 is the angle between the directional vector997888119886119894and the
positive direction of the 119909-axis and 120574 is the length of the 997888119886119894
The registration parameters of the rigid transformderivedabovewere then applied to theMR volumes tomatchwith theSPECT images
23 ROI Segmentation from MR T1-Weighted Images Theregistered MR T1-weighted images obtained from the pre-vious stage were then used as the reference to demarcatethe striatum on the corresponding SPECT images As theassessment of TRODAT-1 BR is usually carried out from theaxial view of SPECT images the segmentation of the striatumwas performed in the axial planes of the MR images Figure 2shows the structure of the striatum from the axial view of theMR image It can be seen that the left and right sides of stria-tum of the basal ganglia are located beside the ventricle Eachside of the striatum can be further divided into the caudatenucleus and putamen We named the two pairs of caudatenucleus and putamen the ST regions However the divisionof the ST regions is not obvious because they usually fuse withother brain structures The unclear cut between the caudatenucleus and putamen and the surrounding brain structurebrings up difficulties in isolating the ST regions solely usingautomatic image segmentation techniqueswithout any expertintervention
To segment these four ROIs we adopted a modifiedactive contour model In this way an initial contour of thefirst slice can be determined by an expert according to thetopological andmorphological characteristics of the STOncethe location and shape of the ST regions are confined into thebond of the initial contour then refinement can be carriedout by the computer according to the intensity informationIn addition assuming smooth variation of the 3D ST regioncontour the final contour of the present slice can be directlyused as the initial contour for the next slice To achieve thisgoal the active contour model is a suitable choice
The basis of the active contour model named snake isto represent an initial contour in the parametric form of
V(119904) = [119909(119904) 119910(119904)] 119904 isin [0 1] that deforms to the optimalshape by minimizing the energy functional
119864snake = int1
0
[119864int (V (119904)) + 119864ext (V (119904))] 119889119904
where 120572 and 120573 are the parameters to weight the influence onthe curve deformation from the curversquos tension V1015840(119904) and therigidity V10158401015840(119904) respectively
Theoretically at the minima of the energy functional thesnake must satisfy the Euler equation
120572V10158401015840(119904) minus 120573V
1015840101584010158401015840(119904) minus nabla119864 ext(V (119904)) = 0 (3)
As the first derivative of energy gives the force the aboveequation can be interpreted as a force balance equation
119865int (V) + 119865ext (V) = 0 (4)
The internal force 119865int(V) = 120572V10158401015840(119904)minus120573V1015840101584010158401015840(119904) restricts the
curve to stretch and bend while the external force 119865ext(V) =minusnabla119864 ext(V) pulls the curve toward the desired image edges
The snake is an active rather than a salient model dueto the dynamic deformation process by treating the forcebalance equation as function of time 119905Therefore the solutionof (3) can be approximated by iteratively searching for thesteady state of the following equation where the V(119904 119905) =[119909(119904 119905) 119910(119904 119905)] denotes V(119904) at the 119905th iteration
120597V (119904 119905)
120597119905
= 120572V10158401015840(119904 119905) minus 120573V
1015840101584010158401015840(119904 119905) minus nabla119864 ext(V (119904 119905)) (5)
In practice a numerical solution to (5) can be achieved bydiscretizing 119904 iteratively using a finite difference method [16]as per
x119905= (A + 120574I)minus1 (x
119905minus1minus p119909119905minus1)
y119905= (A + 120574I)minus1 (y
119905minus1minus p119910119905minus1)
(6)
where A is a pentadiagonal matrix containing the constants120572 and 120573 The parameter of 120574 is the step size to control thedegree of the contour deformation between iterations I isthe unit matrix x
119905and y119905are the vectors consisting of the 119909-
and 119910-coordinates of the contour V(119904 119905) respectively p119909119905minus1
and p119910119905minus1
are the vectors containing 120597119864ext(119909(119904 119905 minus 1) 119910(119904 119905 minus1))120597119909 and 120597119864ext(119909(119904 119905 minus 1) 119910(119904 119905 minus 1))120597119910 as their elementsfor all 119904 respectively
The external force (nabla119864ext) in the active model can usuallybe classified into two types static and dynamic Static forcesare derived from the image gradients which do not changethroughout the deformation process while dynamic forcesvary as the snake deforms Using the image gradient as theexternal force makes the conventional snake difficult to moveinto a concave edge because the null image gradients withina homogenous region inside the contour fail to attract thecontour and as a result the contour is only affected by
Computational and Mathematical Methods in Medicine 5
the internal forces Even though several dynamic externalforces have been proposed to alleviate such a limitation ofthe static external forces they also raised other problemsincreasing the calculation complexity or leading to uncon-trollable deformation [17 18] A new static external forcecalled gradient-vector flow (GVF) adding the directionalproperty into the original image gradient map was proposedby Xu and Prince to improve the performance of the staticsnake in concave edge detection [19] Several reports havedemonstrated the success of applying the GVF snake tomedical image segmentation [20ndash23] including brain MRI[24] This encouraged us to apply the GVF to segment the STregions in our study
The gradient-vector-flow field is defined as k(119909 119910) =[119906(119909 119910) V(119909 119910)] such that the external energy functionbecomes
where 120583 is a parameter to control the degree of smoothness ofthe gradient-vector-flow field and nabla119891 is an edge map derivedfrom the original image 119891(119909 119910)
To solve the equation numerically by discretization anditeration let 119899 indicate the times of iteration and theincrements in 119909 119910 and 119905 are all equal to 1 The relation ofvector flows from the current to the next position can bederived as
There are 4 parameters determined empirically to obtainthe optimal results in the approach The elasticity parameter(120572) and the rigidity parameter (120573) in (2) were set to be 01 and02 respectively The parameter (120574) in (6) was set to be 1 Theexternal force weight (120583) in (7) was set to be 05
3 Results and Discussion
31 Registration Results An example is given in Figure 3 toillustrate the use of our developed interface to detect the slicescontaining the MCSA from the CT and MR volumes andregister these two images through the GHT Once the rigidtransform with the registration parameters had been appliedto the MR image volume it can directly match the SPECTvolume as shown in Figure 4 The registration accuracyreached 9648 Our previous study quantitatively evaluated
Figure 3 An example of registration of the CT image (middle smallpanel) and theMR image (right small panel) to render the final fusedimage (left large panel)The red and blue lines in the small panels arethe detected boundaries and the contour of the head
Table 1 Correspondence of the manual delineation between theobservers
JI () Rater A-Rater B Rater B-Rater C Rater A-Rater CCase 1 604 plusmn 48 621 plusmn 11 756 plusmn 21
Case 2 543 plusmn 45 547 plusmn 56 718 plusmn 52
Case 3 580 plusmn 47 581 plusmn 54 766 plusmn 52
the registration accuracy of the proposed method better thanthe results obtained solely using the PAR method or directlyregistering SPECT with MR images [12]
32 Segmentation Results The expert delineation and theGVF segmentation of the ST regions containing two pairsof the caudate nucleus and putamen are given in Figure 5 Aquantitative comparison of these twomethods is given below
We used the Jaccard index (JI) to quantify the degree ofmatch between the two corresponding ROIsThe JI is definedas the ratio of the intersection of two volumes Ω
1and Ω
2by
the union of them If the two volumes completely overlap theJI value is equal to 100
JI =1003816100381610038161003816Ω1cap Ω2
1003816100381610038161003816
1003816100381610038161003816Ω1cup Ω2
1003816100381610038161003816
times 100 (9)
The five sequential axial slices containing the ST fromthree normal cases were recruited in the comparative eval-uation Three neurologists first manually delineated the STregions including the caudate nucleus and putamen on twolateral sides of the brain The intrarater correspondences interms of the mean and standard deviation of the JI valuesfrom the five slices are listed in Table 1 suggesting greatdifferences between observers It was found that Raters A andChad the highest correspondence with a JI value greater than70
The JI values were also derived by mapping the manuallydefined contours by each rater into the GVF segmentedresults The initial contour of the first slice in each case was
6 Computational and Mathematical Methods in Medicine
(a) (b) (c)
Figure 4 Registration between (a) SPECT and (b) MR to give the final overlaid image in (c)
(a) (b) (c)
Figure 5 The ST regions in (a) the original MR T1-weighted image and the corresponding segmentation results by (b) manual delineationand (c) the GVF snake
Table 2 Correspondence betweenmanual delineation and theGVFsnake result for each observer
JI () Rater A-GVFsnake
Rater B-GVFsnake
Rater C-GVFsnake
Case 1 644 plusmn 90 562 plusmn 55 653 plusmn 48
Case 2 686 plusmn 16 577 plusmn 51 654 plusmn 65
Case 3 611 plusmn 38 515 plusmn 31 599 plusmn 37
defined by the same specialist in the GVF snake processTable 2 shows the correspondence with GVF segmentationWe used the paired 119905-test to evaluate the significance levelbetween the JIs derived from the interrater comparison andthose from the rater-GVF comparison for each slice in eachcase The insignificant differences (119875 = 0124 under a 95confidence interval) suggest that the segmentation accuracyusing the GVF snake is compatible with the level of manualdrawing
Rater A the chief neurologist was required to conductthe manual drawing twice The JI values of the two delin-eations are listed in the middle column of Table 3 showing
Table 3 Correspondence between two repeated conductions ofeach method
that the correspondence declined along with the slice num-ber Instead of segmentation solely by hand the GVF snakewas also applied twice to the same set of images Only thefirst slice required an initial contour manually defined by therater each time The JI values of repeated conduction of theGVF snake are also listed in the third column of Table 3suggesting more stable results than those from slice-by-slicemanual drawing
Computational and Mathematical Methods in Medicine 7
In comparison with the index of overlap (similar to JI)between hand-drawing and computer-aided segmentationreported in the literature [9ndash11] the JI values obtained inour study were relatively low This could be due to the extraregion that is the putamen involved in our study Thesegmentation target of the previous reports is focused onthe caudate nucleuses that are next to ventricles with greatercontrast (Figure 1) and therefore more easily identified Incomparison with the caudate nucleuses the low contrastof the putamen to the surrounding tissue increases thedifficulty of extraction Using expert hand drawing as thecomparison basis seems to be the only choice in currentstudies since there is no gold standard to determine theabsolute accuracy of segmentation of the ST regions dueto individual-dependent variation in the brain structureHowever we demonstrated that significant interobserver andintraobserver variability in such a decision exists even amongthe well-trained neurologists participating in our studywhich was overlooked in previous studies The inconsistencyin decision-making could be incurred by the small size (inthe order of 100 pixels) of the structures as compared with theimaging resolution and image noise Under a compatible levelof precision as shown in Tables 1 and 2 we demonstrated thatthe reproducibility and consistency improved when using theGVF snake segmentationmethod In addition to stability theGVF snake can save labor and provide a more efficient waythan previous studies to define the ST regions contours inconsecutive slices as it only requires an initial contour drawnby hand in the first slice
33 Binding Ratio Calculation In the final stage of evaluationof the reliability of the proposedmethod after completing theMR and SPECT image registration the BRs were also derivedfrom the segmented ST regions in the SPECT images usingthe proposed method to compare with those obtained usingcommercial software (Siemens Medical Systems KnoxvilleTN USA) in which the ST regions were manually outlinedby an expert The BR was calculated by normalizing themean intensity in the ST regions by the mean intensity inthe occipital cortices Linear regression analysis (Figure 6)revealed a close correlation (CC = 0874 under 95confidence interval) between the BRs derived by the twosystems
4 Conclusions
To calculate the regional TRODAT-1 binding ratio in SPECTstudies accurate and repeatable extraction of the ST regionsfrom MR images is required to indirectly define the cor-responding regions in the SPECT images Segmentationdirectly on the SPECT image is not applicable in this casebecause it distorts the ST regions Clinical routine tends toapply manual delineation of the ST regions which is proneto errors incurred through interobserver and intraobservervariability Previous researchers have developed several seg-mentation algorithms to complete similar tasks where expertdecisions for anatomical and morphological informationwere still necessary to refine the results As the localization
1
11
12
13
14
15
16
1 12 14 16 18 2 22 24 26
Our
met
hod
Commercial software
y = 0466x + 0415
R2= 0764
Correlation coefficient = 0874
Figure 6 Linear regression analysis between the BRs derived by ourmethod and those by the commercial software
of the ST regions is a knowledge-driven task the proposedmethod allowed the expert to assign the initial contourin the proper location and applied the gradient-vector-flow snake to approach the real contours In such a waythe complexity of the algorithm can be reduced and theefficiency of segmentation can be increased Results fromthree normal subjects showed a higher reproducibility ofthe proposed method than manual segmentation undercompatible segmentation accuracy The MR images withsegmented ST regions were overlaid on the SPECT imagesusing our previously developed registration algorithm tocalculate the TRODAT-1 BRThe derived BRs correlated wellwith those derived using commercial software suggesting agood reliability of the proposed method
Acknowledgments
The authors would like to thank Dr Chen Nai-Ching andDr Chi-Wei Huang from the Department of NeurologyChang Gung Memorial Hospital Kaohsiung Medical CenterKaohsiung Taiwan for their kind assistance in the task ofmanual drawing of the ST regions in this study
References
[1] S Asenbaum ldquoNuclear medicine in neurodegenerative disor-dersrdquo Imaging Decisions MRI vol 6 pp 19ndash28 2002
[2] K L Chou H I Hurtig M B Stern et al ldquoDiagnostic accuracyof [ 99mTc]TRODAT-1 SPECT imaging in early Parkinsonrsquosdiseaserdquo Parkinsonism and Related Disorders vol 10 no 6 pp375ndash379 2004
[3] J L Cummings C Henchcliffe S Schaier T Simuni AWaxman and P Kemp ldquoThe role of dopaminergic imaging inpatients with symptoms of dopaminergic system neurodegen-erationrdquo Brain vol 134 no 11 pp 3146ndash3166 2011
[4] A Siderowf A Newberg K L Chou et al ldquo[ 99mTc]TRODAT-1 SPECT imaging correlates with odor identification in earlyParkinson diseaserdquo Neurology vol 64 no 10 pp 1716ndash17202005
8 Computational and Mathematical Methods in Medicine
[5] Y-H Weng T-C Yen M-C Chen et al ldquoSensitivity andspecificity of 99mTc-TRODAT-1 SPECT imaging in differenti-ating patients with idiopathic Parkinsonrsquos disease from healthysubjectsrdquo Journal of NuclearMedicine vol 45 no 3 pp 393ndash4012004
[6] W-S Huang S-Z Lin J-C Lin S-P Wey G Ting and R-SLiu ldquoEvaluation of early-stage Parkinsonrsquos disease with 99mTc-TRODAT-1 imagingrdquo Journal of NuclearMedicine vol 42 no 9pp 1303ndash1308 2001
[7] P D Acton P T Meyer P D Mozley K Plossl and H F KungldquoSimplified quantification of dopamine transporters in humansusing [ 99mTc]TRODAT-1 and single-photon emission tomog-raphyrdquo European Journal of Nuclear Medicine and MolecularImaging vol 27 no 11 pp 1714ndash1718 2000
[8] P D Acton P D Mozley and H F Kung ldquoLogistic discrimi-nant parametric mapping a novel method for the pixel-baseddifferential diagnosis of Parkinsonrsquos diseaserdquo European Journalof Nuclear Medicine and Molecular Imaging vol 26 no 11 pp1413ndash1423 1999
[9] A J Worth N Makris M R Patti et al ldquoPrecise segmentationof the lateral ventricles and caudate nucleus in mr brain imagesusing anatomically driven histogramsrdquo IEEE Transactions onMedical Imaging vol 17 no 2 pp 303ndash310 1998
[10] V Barra and J-Y Boire ldquoAutomatic segmentation of subcorticalbrain structures in MR images using information fusionrdquo IEEETransactions on Medical Imaging vol 20 no 7 pp 549ndash5582001
[11] Y Xia K Bettinger L Shen and A L Reiss ldquoAutomaticsegmentation of the caudate nucleus from human brain MRimagesrdquo IEEE Transactions on Medical Imaging vol 26 no 4pp 509ndash517 2007
[12] C-F Jiang C-C Chang and S-H Huang ldquoRegions of interestextraction from SPECT images for neural degeneration assess-ment using multimodality image fusionrdquo MultidimensionalSystems and Signal Processing vol 23 pp 437ndash449 2012
[13] A P Dhawan ldquoImage registrationrdquo in Medical Image AnalysisM Akay Ed pp 251ndash276 IEEE Press Piscataway NJ USA2003
[14] A P Dhawan L K Arata A V Levy and J Mantil ldquoIterativeprincipal axes registrationmethod for analysis ofMR-PETbrainimagesrdquo IEEE Transactions on Biomedical Engineering vol 42no 11 pp 1079ndash1087 1995
[15] C-F Jiang C-H Huang and S-T Yang ldquoUsingmaximal cross-section detection for the registration of 3D image data of theheadrdquo Journal of Medical and Biological Engineering vol 31 no3 pp 217ndash226 2011
[16] MKass AWitkin andD Terzopoulos ldquoSnakes active contourmodelsrdquo International Journal of Computer Vision vol 1 no 4pp 321ndash331 1988
[17] L D Cohen and I Cohen ldquoFinite-element methods for activecontour models and balloons for 2-D and 3-D imagesrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol15 no 11 pp 1131ndash1147 1993
[18] B Leroy I Herlin and L Cohen ldquoMulti-resolution algorithmsfor active contour modelsrdquo in Proceedings of the 12th Inter-national Conference on Analysis and Optimization of SystemsImages (ICAOS rsquo96) pp 58ndash65 1996
[19] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998
[20] B Erkol R H Moss R J Stanley W V Stoecker and EHvatum ldquoAutomatic lesion boundary detection in dermoscopy
images using gradient vector flow snakesrdquo Skin Research andTechnology vol 11 no 1 pp 17ndash26 2005
[21] J Tang and S T Acton ldquoVessel boundary tracking for intravitalmicroscopy via multiscale gradient vector flow snakesrdquo IEEETransactions on Biomedical Engineering vol 51 no 2 pp 316ndash324 2004
[22] J Tang S Millington S T Acton J Crandall and S HurwitzldquoSurface extraction and thickness measurement of the articularcartilage fromMR images using directional gradient vector flowsnakesrdquo IEEE Transactions on Biomedical Engineering vol 53no 5 pp 896ndash907 2006
[23] N Tanki K Murase M Kumashiro et al ldquoQuantification ofleft ventricular volumes from cardiac cine MRI using activecontour model combined with gradient vector flowrdquo MagneticResonance in Medical Sciences vol 4 no 4 pp 191ndash196 2005
[24] C Xu and J L Prince ldquoGradient vector flow deformable mod-elsrdquo in Handbook of Medical Imaging pp 159ndash169 AcademicPress 2000
Computational and Mathematical Methods in Medicine 3
Vertical axes alignment
Voxel size adjustment
Derivation of the principle axes
Generalized Hough transform
ROI segmentation
Fine-tune
Coarse registration
Original images
Detect MCSA
Search for theoptimal parameters
MR imagesHead segmentationCT images
SPECT images
SPECT CT scan
The active contour model gradient-vector flow
Rotate and shift
Map MR ST regions into SPECTto calculate binding ratio
Figure 1 The overall process to derive the BRs in SPECT images via registration of the images from SPECT-CT and MR with ROIsegmentation from the registered MR images
registration (PAR) or mutual information (MI) To alleviatethis problem we developed a hybrid registration methodcombining principal axes registrationwith the generalHoughtransform [12] In addition we took advantage of SPECT-CTwhich can acquire SPECT andCT images simultaneouslywhile the patient maintains hisher position on the samecouch Registration used the CT image as the registrationmedium to increase the registration accuracy between theSPECT and the MR image volumes The registration processis fully automatic The essential idea of the design is brieflydescribed below
The voxel size was adjusted to a 1mm3 cube throughbicubic interpolation prior to the following registrationprocess The 3D head was segmented as an entity to deriveits three principal axes prior to registration In this two-stage registration scheme principal axes registration was firstapplied for coarse registration followed by fine-tuning viaapplying the general Hough transform to the contour of themaximal cross-sectional area (MCSA) The original conceptof principal axes registration (PAR) is to superimpose thetwo volumes by aligning the corresponding three principalaxes from both head volumes [13] However the registrationaccuracy of PAR is restricted by the degree of correspondencebetween the two sets of principal axes [14] As the scanningrange of one image modality is usually not the same as theother the centroids of the two different volume sets wouldnot be identical In consequence the two sets of principalaxes derived from the different centroids do not coincidewith each other Therefore in the coarse-registration stage
we only adopted PAR to adjust the orientations of the longaxis of the head to be parallel with the 119911-axis of the systemcoordinates After this stage the long axis from both headvolumes coincided with each other but the horizontal planeswith the two short axes from the two volumes were stillmismatched
In the second stage the registration error in the hori-zontal plane was then fine-tuned We then turned the 3Dregistration task into a 2D one by searching for the slicescontaining the (MCSA) in both volumes in that we hadproved the reliability of using the MCSA as the anatomicalfeature for registration [15] The vertical shift was firstcorrected by aligning these two slices then the detectedcontour of the MCSA was used to derive the registrationparameters via the generalized Hough transform (GHT)Theprocess of the GHT algorithm in this approach includedtwo steps First an R-table was built by calculating thevector set 997888119886
119894 between each contour point (119909
119894 119910119894) and
the center of the contour 119875119888(119909119888 119910119888) in the CT image
Then the corresponding center point (1198751015840119888) was derived by
searching for the maximal intersection via remapping thevector information to each contour point X
119894(119909119894 119910119894) in the
MR image In this study as there was no scale for referenceand the voxel size had been adjusted to be the same weadapted a robust search only for the rotation angle 120573 in (1)when the optimal match between 119875
119888and 1198751015840
119888was achieved
119909119888= 119909119894+ 120574 cos (120579 + 120573)
119910119888= 119910119894+ 120574 sin (120579 + 120573)
(1)
4 Computational and Mathematical Methods in Medicine
Caudate nucleus
Putamen
Left striatum of basal ganglia
Lateral ventricle
Figure 2 Anatomical structure of the striatum from the axial viewof an MR T1-weighted image
where 120579 is the angle between the directional vector997888119886119894and the
positive direction of the 119909-axis and 120574 is the length of the 997888119886119894
The registration parameters of the rigid transformderivedabovewere then applied to theMR volumes tomatchwith theSPECT images
23 ROI Segmentation from MR T1-Weighted Images Theregistered MR T1-weighted images obtained from the pre-vious stage were then used as the reference to demarcatethe striatum on the corresponding SPECT images As theassessment of TRODAT-1 BR is usually carried out from theaxial view of SPECT images the segmentation of the striatumwas performed in the axial planes of the MR images Figure 2shows the structure of the striatum from the axial view of theMR image It can be seen that the left and right sides of stria-tum of the basal ganglia are located beside the ventricle Eachside of the striatum can be further divided into the caudatenucleus and putamen We named the two pairs of caudatenucleus and putamen the ST regions However the divisionof the ST regions is not obvious because they usually fuse withother brain structures The unclear cut between the caudatenucleus and putamen and the surrounding brain structurebrings up difficulties in isolating the ST regions solely usingautomatic image segmentation techniqueswithout any expertintervention
To segment these four ROIs we adopted a modifiedactive contour model In this way an initial contour of thefirst slice can be determined by an expert according to thetopological andmorphological characteristics of the STOncethe location and shape of the ST regions are confined into thebond of the initial contour then refinement can be carriedout by the computer according to the intensity informationIn addition assuming smooth variation of the 3D ST regioncontour the final contour of the present slice can be directlyused as the initial contour for the next slice To achieve thisgoal the active contour model is a suitable choice
The basis of the active contour model named snake isto represent an initial contour in the parametric form of
V(119904) = [119909(119904) 119910(119904)] 119904 isin [0 1] that deforms to the optimalshape by minimizing the energy functional
119864snake = int1
0
[119864int (V (119904)) + 119864ext (V (119904))] 119889119904
where 120572 and 120573 are the parameters to weight the influence onthe curve deformation from the curversquos tension V1015840(119904) and therigidity V10158401015840(119904) respectively
Theoretically at the minima of the energy functional thesnake must satisfy the Euler equation
120572V10158401015840(119904) minus 120573V
1015840101584010158401015840(119904) minus nabla119864 ext(V (119904)) = 0 (3)
As the first derivative of energy gives the force the aboveequation can be interpreted as a force balance equation
119865int (V) + 119865ext (V) = 0 (4)
The internal force 119865int(V) = 120572V10158401015840(119904)minus120573V1015840101584010158401015840(119904) restricts the
curve to stretch and bend while the external force 119865ext(V) =minusnabla119864 ext(V) pulls the curve toward the desired image edges
The snake is an active rather than a salient model dueto the dynamic deformation process by treating the forcebalance equation as function of time 119905Therefore the solutionof (3) can be approximated by iteratively searching for thesteady state of the following equation where the V(119904 119905) =[119909(119904 119905) 119910(119904 119905)] denotes V(119904) at the 119905th iteration
120597V (119904 119905)
120597119905
= 120572V10158401015840(119904 119905) minus 120573V
1015840101584010158401015840(119904 119905) minus nabla119864 ext(V (119904 119905)) (5)
In practice a numerical solution to (5) can be achieved bydiscretizing 119904 iteratively using a finite difference method [16]as per
x119905= (A + 120574I)minus1 (x
119905minus1minus p119909119905minus1)
y119905= (A + 120574I)minus1 (y
119905minus1minus p119910119905minus1)
(6)
where A is a pentadiagonal matrix containing the constants120572 and 120573 The parameter of 120574 is the step size to control thedegree of the contour deformation between iterations I isthe unit matrix x
119905and y119905are the vectors consisting of the 119909-
and 119910-coordinates of the contour V(119904 119905) respectively p119909119905minus1
and p119910119905minus1
are the vectors containing 120597119864ext(119909(119904 119905 minus 1) 119910(119904 119905 minus1))120597119909 and 120597119864ext(119909(119904 119905 minus 1) 119910(119904 119905 minus 1))120597119910 as their elementsfor all 119904 respectively
The external force (nabla119864ext) in the active model can usuallybe classified into two types static and dynamic Static forcesare derived from the image gradients which do not changethroughout the deformation process while dynamic forcesvary as the snake deforms Using the image gradient as theexternal force makes the conventional snake difficult to moveinto a concave edge because the null image gradients withina homogenous region inside the contour fail to attract thecontour and as a result the contour is only affected by
Computational and Mathematical Methods in Medicine 5
the internal forces Even though several dynamic externalforces have been proposed to alleviate such a limitation ofthe static external forces they also raised other problemsincreasing the calculation complexity or leading to uncon-trollable deformation [17 18] A new static external forcecalled gradient-vector flow (GVF) adding the directionalproperty into the original image gradient map was proposedby Xu and Prince to improve the performance of the staticsnake in concave edge detection [19] Several reports havedemonstrated the success of applying the GVF snake tomedical image segmentation [20ndash23] including brain MRI[24] This encouraged us to apply the GVF to segment the STregions in our study
The gradient-vector-flow field is defined as k(119909 119910) =[119906(119909 119910) V(119909 119910)] such that the external energy functionbecomes
where 120583 is a parameter to control the degree of smoothness ofthe gradient-vector-flow field and nabla119891 is an edge map derivedfrom the original image 119891(119909 119910)
To solve the equation numerically by discretization anditeration let 119899 indicate the times of iteration and theincrements in 119909 119910 and 119905 are all equal to 1 The relation ofvector flows from the current to the next position can bederived as
There are 4 parameters determined empirically to obtainthe optimal results in the approach The elasticity parameter(120572) and the rigidity parameter (120573) in (2) were set to be 01 and02 respectively The parameter (120574) in (6) was set to be 1 Theexternal force weight (120583) in (7) was set to be 05
3 Results and Discussion
31 Registration Results An example is given in Figure 3 toillustrate the use of our developed interface to detect the slicescontaining the MCSA from the CT and MR volumes andregister these two images through the GHT Once the rigidtransform with the registration parameters had been appliedto the MR image volume it can directly match the SPECTvolume as shown in Figure 4 The registration accuracyreached 9648 Our previous study quantitatively evaluated
Figure 3 An example of registration of the CT image (middle smallpanel) and theMR image (right small panel) to render the final fusedimage (left large panel)The red and blue lines in the small panels arethe detected boundaries and the contour of the head
Table 1 Correspondence of the manual delineation between theobservers
JI () Rater A-Rater B Rater B-Rater C Rater A-Rater CCase 1 604 plusmn 48 621 plusmn 11 756 plusmn 21
Case 2 543 plusmn 45 547 plusmn 56 718 plusmn 52
Case 3 580 plusmn 47 581 plusmn 54 766 plusmn 52
the registration accuracy of the proposed method better thanthe results obtained solely using the PAR method or directlyregistering SPECT with MR images [12]
32 Segmentation Results The expert delineation and theGVF segmentation of the ST regions containing two pairsof the caudate nucleus and putamen are given in Figure 5 Aquantitative comparison of these twomethods is given below
We used the Jaccard index (JI) to quantify the degree ofmatch between the two corresponding ROIsThe JI is definedas the ratio of the intersection of two volumes Ω
1and Ω
2by
the union of them If the two volumes completely overlap theJI value is equal to 100
JI =1003816100381610038161003816Ω1cap Ω2
1003816100381610038161003816
1003816100381610038161003816Ω1cup Ω2
1003816100381610038161003816
times 100 (9)
The five sequential axial slices containing the ST fromthree normal cases were recruited in the comparative eval-uation Three neurologists first manually delineated the STregions including the caudate nucleus and putamen on twolateral sides of the brain The intrarater correspondences interms of the mean and standard deviation of the JI valuesfrom the five slices are listed in Table 1 suggesting greatdifferences between observers It was found that Raters A andChad the highest correspondence with a JI value greater than70
The JI values were also derived by mapping the manuallydefined contours by each rater into the GVF segmentedresults The initial contour of the first slice in each case was
6 Computational and Mathematical Methods in Medicine
(a) (b) (c)
Figure 4 Registration between (a) SPECT and (b) MR to give the final overlaid image in (c)
(a) (b) (c)
Figure 5 The ST regions in (a) the original MR T1-weighted image and the corresponding segmentation results by (b) manual delineationand (c) the GVF snake
Table 2 Correspondence betweenmanual delineation and theGVFsnake result for each observer
JI () Rater A-GVFsnake
Rater B-GVFsnake
Rater C-GVFsnake
Case 1 644 plusmn 90 562 plusmn 55 653 plusmn 48
Case 2 686 plusmn 16 577 plusmn 51 654 plusmn 65
Case 3 611 plusmn 38 515 plusmn 31 599 plusmn 37
defined by the same specialist in the GVF snake processTable 2 shows the correspondence with GVF segmentationWe used the paired 119905-test to evaluate the significance levelbetween the JIs derived from the interrater comparison andthose from the rater-GVF comparison for each slice in eachcase The insignificant differences (119875 = 0124 under a 95confidence interval) suggest that the segmentation accuracyusing the GVF snake is compatible with the level of manualdrawing
Rater A the chief neurologist was required to conductthe manual drawing twice The JI values of the two delin-eations are listed in the middle column of Table 3 showing
Table 3 Correspondence between two repeated conductions ofeach method
that the correspondence declined along with the slice num-ber Instead of segmentation solely by hand the GVF snakewas also applied twice to the same set of images Only thefirst slice required an initial contour manually defined by therater each time The JI values of repeated conduction of theGVF snake are also listed in the third column of Table 3suggesting more stable results than those from slice-by-slicemanual drawing
Computational and Mathematical Methods in Medicine 7
In comparison with the index of overlap (similar to JI)between hand-drawing and computer-aided segmentationreported in the literature [9ndash11] the JI values obtained inour study were relatively low This could be due to the extraregion that is the putamen involved in our study Thesegmentation target of the previous reports is focused onthe caudate nucleuses that are next to ventricles with greatercontrast (Figure 1) and therefore more easily identified Incomparison with the caudate nucleuses the low contrastof the putamen to the surrounding tissue increases thedifficulty of extraction Using expert hand drawing as thecomparison basis seems to be the only choice in currentstudies since there is no gold standard to determine theabsolute accuracy of segmentation of the ST regions dueto individual-dependent variation in the brain structureHowever we demonstrated that significant interobserver andintraobserver variability in such a decision exists even amongthe well-trained neurologists participating in our studywhich was overlooked in previous studies The inconsistencyin decision-making could be incurred by the small size (inthe order of 100 pixels) of the structures as compared with theimaging resolution and image noise Under a compatible levelof precision as shown in Tables 1 and 2 we demonstrated thatthe reproducibility and consistency improved when using theGVF snake segmentationmethod In addition to stability theGVF snake can save labor and provide a more efficient waythan previous studies to define the ST regions contours inconsecutive slices as it only requires an initial contour drawnby hand in the first slice
33 Binding Ratio Calculation In the final stage of evaluationof the reliability of the proposedmethod after completing theMR and SPECT image registration the BRs were also derivedfrom the segmented ST regions in the SPECT images usingthe proposed method to compare with those obtained usingcommercial software (Siemens Medical Systems KnoxvilleTN USA) in which the ST regions were manually outlinedby an expert The BR was calculated by normalizing themean intensity in the ST regions by the mean intensity inthe occipital cortices Linear regression analysis (Figure 6)revealed a close correlation (CC = 0874 under 95confidence interval) between the BRs derived by the twosystems
4 Conclusions
To calculate the regional TRODAT-1 binding ratio in SPECTstudies accurate and repeatable extraction of the ST regionsfrom MR images is required to indirectly define the cor-responding regions in the SPECT images Segmentationdirectly on the SPECT image is not applicable in this casebecause it distorts the ST regions Clinical routine tends toapply manual delineation of the ST regions which is proneto errors incurred through interobserver and intraobservervariability Previous researchers have developed several seg-mentation algorithms to complete similar tasks where expertdecisions for anatomical and morphological informationwere still necessary to refine the results As the localization
1
11
12
13
14
15
16
1 12 14 16 18 2 22 24 26
Our
met
hod
Commercial software
y = 0466x + 0415
R2= 0764
Correlation coefficient = 0874
Figure 6 Linear regression analysis between the BRs derived by ourmethod and those by the commercial software
of the ST regions is a knowledge-driven task the proposedmethod allowed the expert to assign the initial contourin the proper location and applied the gradient-vector-flow snake to approach the real contours In such a waythe complexity of the algorithm can be reduced and theefficiency of segmentation can be increased Results fromthree normal subjects showed a higher reproducibility ofthe proposed method than manual segmentation undercompatible segmentation accuracy The MR images withsegmented ST regions were overlaid on the SPECT imagesusing our previously developed registration algorithm tocalculate the TRODAT-1 BRThe derived BRs correlated wellwith those derived using commercial software suggesting agood reliability of the proposed method
Acknowledgments
The authors would like to thank Dr Chen Nai-Ching andDr Chi-Wei Huang from the Department of NeurologyChang Gung Memorial Hospital Kaohsiung Medical CenterKaohsiung Taiwan for their kind assistance in the task ofmanual drawing of the ST regions in this study
References
[1] S Asenbaum ldquoNuclear medicine in neurodegenerative disor-dersrdquo Imaging Decisions MRI vol 6 pp 19ndash28 2002
[2] K L Chou H I Hurtig M B Stern et al ldquoDiagnostic accuracyof [ 99mTc]TRODAT-1 SPECT imaging in early Parkinsonrsquosdiseaserdquo Parkinsonism and Related Disorders vol 10 no 6 pp375ndash379 2004
[3] J L Cummings C Henchcliffe S Schaier T Simuni AWaxman and P Kemp ldquoThe role of dopaminergic imaging inpatients with symptoms of dopaminergic system neurodegen-erationrdquo Brain vol 134 no 11 pp 3146ndash3166 2011
[4] A Siderowf A Newberg K L Chou et al ldquo[ 99mTc]TRODAT-1 SPECT imaging correlates with odor identification in earlyParkinson diseaserdquo Neurology vol 64 no 10 pp 1716ndash17202005
8 Computational and Mathematical Methods in Medicine
[5] Y-H Weng T-C Yen M-C Chen et al ldquoSensitivity andspecificity of 99mTc-TRODAT-1 SPECT imaging in differenti-ating patients with idiopathic Parkinsonrsquos disease from healthysubjectsrdquo Journal of NuclearMedicine vol 45 no 3 pp 393ndash4012004
[6] W-S Huang S-Z Lin J-C Lin S-P Wey G Ting and R-SLiu ldquoEvaluation of early-stage Parkinsonrsquos disease with 99mTc-TRODAT-1 imagingrdquo Journal of NuclearMedicine vol 42 no 9pp 1303ndash1308 2001
[7] P D Acton P T Meyer P D Mozley K Plossl and H F KungldquoSimplified quantification of dopamine transporters in humansusing [ 99mTc]TRODAT-1 and single-photon emission tomog-raphyrdquo European Journal of Nuclear Medicine and MolecularImaging vol 27 no 11 pp 1714ndash1718 2000
[8] P D Acton P D Mozley and H F Kung ldquoLogistic discrimi-nant parametric mapping a novel method for the pixel-baseddifferential diagnosis of Parkinsonrsquos diseaserdquo European Journalof Nuclear Medicine and Molecular Imaging vol 26 no 11 pp1413ndash1423 1999
[9] A J Worth N Makris M R Patti et al ldquoPrecise segmentationof the lateral ventricles and caudate nucleus in mr brain imagesusing anatomically driven histogramsrdquo IEEE Transactions onMedical Imaging vol 17 no 2 pp 303ndash310 1998
[10] V Barra and J-Y Boire ldquoAutomatic segmentation of subcorticalbrain structures in MR images using information fusionrdquo IEEETransactions on Medical Imaging vol 20 no 7 pp 549ndash5582001
[11] Y Xia K Bettinger L Shen and A L Reiss ldquoAutomaticsegmentation of the caudate nucleus from human brain MRimagesrdquo IEEE Transactions on Medical Imaging vol 26 no 4pp 509ndash517 2007
[12] C-F Jiang C-C Chang and S-H Huang ldquoRegions of interestextraction from SPECT images for neural degeneration assess-ment using multimodality image fusionrdquo MultidimensionalSystems and Signal Processing vol 23 pp 437ndash449 2012
[13] A P Dhawan ldquoImage registrationrdquo in Medical Image AnalysisM Akay Ed pp 251ndash276 IEEE Press Piscataway NJ USA2003
[14] A P Dhawan L K Arata A V Levy and J Mantil ldquoIterativeprincipal axes registrationmethod for analysis ofMR-PETbrainimagesrdquo IEEE Transactions on Biomedical Engineering vol 42no 11 pp 1079ndash1087 1995
[15] C-F Jiang C-H Huang and S-T Yang ldquoUsingmaximal cross-section detection for the registration of 3D image data of theheadrdquo Journal of Medical and Biological Engineering vol 31 no3 pp 217ndash226 2011
[16] MKass AWitkin andD Terzopoulos ldquoSnakes active contourmodelsrdquo International Journal of Computer Vision vol 1 no 4pp 321ndash331 1988
[17] L D Cohen and I Cohen ldquoFinite-element methods for activecontour models and balloons for 2-D and 3-D imagesrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol15 no 11 pp 1131ndash1147 1993
[18] B Leroy I Herlin and L Cohen ldquoMulti-resolution algorithmsfor active contour modelsrdquo in Proceedings of the 12th Inter-national Conference on Analysis and Optimization of SystemsImages (ICAOS rsquo96) pp 58ndash65 1996
[19] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998
[20] B Erkol R H Moss R J Stanley W V Stoecker and EHvatum ldquoAutomatic lesion boundary detection in dermoscopy
images using gradient vector flow snakesrdquo Skin Research andTechnology vol 11 no 1 pp 17ndash26 2005
[21] J Tang and S T Acton ldquoVessel boundary tracking for intravitalmicroscopy via multiscale gradient vector flow snakesrdquo IEEETransactions on Biomedical Engineering vol 51 no 2 pp 316ndash324 2004
[22] J Tang S Millington S T Acton J Crandall and S HurwitzldquoSurface extraction and thickness measurement of the articularcartilage fromMR images using directional gradient vector flowsnakesrdquo IEEE Transactions on Biomedical Engineering vol 53no 5 pp 896ndash907 2006
[23] N Tanki K Murase M Kumashiro et al ldquoQuantification ofleft ventricular volumes from cardiac cine MRI using activecontour model combined with gradient vector flowrdquo MagneticResonance in Medical Sciences vol 4 no 4 pp 191ndash196 2005
[24] C Xu and J L Prince ldquoGradient vector flow deformable mod-elsrdquo in Handbook of Medical Imaging pp 159ndash169 AcademicPress 2000
4 Computational and Mathematical Methods in Medicine
Caudate nucleus
Putamen
Left striatum of basal ganglia
Lateral ventricle
Figure 2 Anatomical structure of the striatum from the axial viewof an MR T1-weighted image
where 120579 is the angle between the directional vector997888119886119894and the
positive direction of the 119909-axis and 120574 is the length of the 997888119886119894
The registration parameters of the rigid transformderivedabovewere then applied to theMR volumes tomatchwith theSPECT images
23 ROI Segmentation from MR T1-Weighted Images Theregistered MR T1-weighted images obtained from the pre-vious stage were then used as the reference to demarcatethe striatum on the corresponding SPECT images As theassessment of TRODAT-1 BR is usually carried out from theaxial view of SPECT images the segmentation of the striatumwas performed in the axial planes of the MR images Figure 2shows the structure of the striatum from the axial view of theMR image It can be seen that the left and right sides of stria-tum of the basal ganglia are located beside the ventricle Eachside of the striatum can be further divided into the caudatenucleus and putamen We named the two pairs of caudatenucleus and putamen the ST regions However the divisionof the ST regions is not obvious because they usually fuse withother brain structures The unclear cut between the caudatenucleus and putamen and the surrounding brain structurebrings up difficulties in isolating the ST regions solely usingautomatic image segmentation techniqueswithout any expertintervention
To segment these four ROIs we adopted a modifiedactive contour model In this way an initial contour of thefirst slice can be determined by an expert according to thetopological andmorphological characteristics of the STOncethe location and shape of the ST regions are confined into thebond of the initial contour then refinement can be carriedout by the computer according to the intensity informationIn addition assuming smooth variation of the 3D ST regioncontour the final contour of the present slice can be directlyused as the initial contour for the next slice To achieve thisgoal the active contour model is a suitable choice
The basis of the active contour model named snake isto represent an initial contour in the parametric form of
V(119904) = [119909(119904) 119910(119904)] 119904 isin [0 1] that deforms to the optimalshape by minimizing the energy functional
119864snake = int1
0
[119864int (V (119904)) + 119864ext (V (119904))] 119889119904
where 120572 and 120573 are the parameters to weight the influence onthe curve deformation from the curversquos tension V1015840(119904) and therigidity V10158401015840(119904) respectively
Theoretically at the minima of the energy functional thesnake must satisfy the Euler equation
120572V10158401015840(119904) minus 120573V
1015840101584010158401015840(119904) minus nabla119864 ext(V (119904)) = 0 (3)
As the first derivative of energy gives the force the aboveequation can be interpreted as a force balance equation
119865int (V) + 119865ext (V) = 0 (4)
The internal force 119865int(V) = 120572V10158401015840(119904)minus120573V1015840101584010158401015840(119904) restricts the
curve to stretch and bend while the external force 119865ext(V) =minusnabla119864 ext(V) pulls the curve toward the desired image edges
The snake is an active rather than a salient model dueto the dynamic deformation process by treating the forcebalance equation as function of time 119905Therefore the solutionof (3) can be approximated by iteratively searching for thesteady state of the following equation where the V(119904 119905) =[119909(119904 119905) 119910(119904 119905)] denotes V(119904) at the 119905th iteration
120597V (119904 119905)
120597119905
= 120572V10158401015840(119904 119905) minus 120573V
1015840101584010158401015840(119904 119905) minus nabla119864 ext(V (119904 119905)) (5)
In practice a numerical solution to (5) can be achieved bydiscretizing 119904 iteratively using a finite difference method [16]as per
x119905= (A + 120574I)minus1 (x
119905minus1minus p119909119905minus1)
y119905= (A + 120574I)minus1 (y
119905minus1minus p119910119905minus1)
(6)
where A is a pentadiagonal matrix containing the constants120572 and 120573 The parameter of 120574 is the step size to control thedegree of the contour deformation between iterations I isthe unit matrix x
119905and y119905are the vectors consisting of the 119909-
and 119910-coordinates of the contour V(119904 119905) respectively p119909119905minus1
and p119910119905minus1
are the vectors containing 120597119864ext(119909(119904 119905 minus 1) 119910(119904 119905 minus1))120597119909 and 120597119864ext(119909(119904 119905 minus 1) 119910(119904 119905 minus 1))120597119910 as their elementsfor all 119904 respectively
The external force (nabla119864ext) in the active model can usuallybe classified into two types static and dynamic Static forcesare derived from the image gradients which do not changethroughout the deformation process while dynamic forcesvary as the snake deforms Using the image gradient as theexternal force makes the conventional snake difficult to moveinto a concave edge because the null image gradients withina homogenous region inside the contour fail to attract thecontour and as a result the contour is only affected by
Computational and Mathematical Methods in Medicine 5
the internal forces Even though several dynamic externalforces have been proposed to alleviate such a limitation ofthe static external forces they also raised other problemsincreasing the calculation complexity or leading to uncon-trollable deformation [17 18] A new static external forcecalled gradient-vector flow (GVF) adding the directionalproperty into the original image gradient map was proposedby Xu and Prince to improve the performance of the staticsnake in concave edge detection [19] Several reports havedemonstrated the success of applying the GVF snake tomedical image segmentation [20ndash23] including brain MRI[24] This encouraged us to apply the GVF to segment the STregions in our study
The gradient-vector-flow field is defined as k(119909 119910) =[119906(119909 119910) V(119909 119910)] such that the external energy functionbecomes
where 120583 is a parameter to control the degree of smoothness ofthe gradient-vector-flow field and nabla119891 is an edge map derivedfrom the original image 119891(119909 119910)
To solve the equation numerically by discretization anditeration let 119899 indicate the times of iteration and theincrements in 119909 119910 and 119905 are all equal to 1 The relation ofvector flows from the current to the next position can bederived as
There are 4 parameters determined empirically to obtainthe optimal results in the approach The elasticity parameter(120572) and the rigidity parameter (120573) in (2) were set to be 01 and02 respectively The parameter (120574) in (6) was set to be 1 Theexternal force weight (120583) in (7) was set to be 05
3 Results and Discussion
31 Registration Results An example is given in Figure 3 toillustrate the use of our developed interface to detect the slicescontaining the MCSA from the CT and MR volumes andregister these two images through the GHT Once the rigidtransform with the registration parameters had been appliedto the MR image volume it can directly match the SPECTvolume as shown in Figure 4 The registration accuracyreached 9648 Our previous study quantitatively evaluated
Figure 3 An example of registration of the CT image (middle smallpanel) and theMR image (right small panel) to render the final fusedimage (left large panel)The red and blue lines in the small panels arethe detected boundaries and the contour of the head
Table 1 Correspondence of the manual delineation between theobservers
JI () Rater A-Rater B Rater B-Rater C Rater A-Rater CCase 1 604 plusmn 48 621 plusmn 11 756 plusmn 21
Case 2 543 plusmn 45 547 plusmn 56 718 plusmn 52
Case 3 580 plusmn 47 581 plusmn 54 766 plusmn 52
the registration accuracy of the proposed method better thanthe results obtained solely using the PAR method or directlyregistering SPECT with MR images [12]
32 Segmentation Results The expert delineation and theGVF segmentation of the ST regions containing two pairsof the caudate nucleus and putamen are given in Figure 5 Aquantitative comparison of these twomethods is given below
We used the Jaccard index (JI) to quantify the degree ofmatch between the two corresponding ROIsThe JI is definedas the ratio of the intersection of two volumes Ω
1and Ω
2by
the union of them If the two volumes completely overlap theJI value is equal to 100
JI =1003816100381610038161003816Ω1cap Ω2
1003816100381610038161003816
1003816100381610038161003816Ω1cup Ω2
1003816100381610038161003816
times 100 (9)
The five sequential axial slices containing the ST fromthree normal cases were recruited in the comparative eval-uation Three neurologists first manually delineated the STregions including the caudate nucleus and putamen on twolateral sides of the brain The intrarater correspondences interms of the mean and standard deviation of the JI valuesfrom the five slices are listed in Table 1 suggesting greatdifferences between observers It was found that Raters A andChad the highest correspondence with a JI value greater than70
The JI values were also derived by mapping the manuallydefined contours by each rater into the GVF segmentedresults The initial contour of the first slice in each case was
6 Computational and Mathematical Methods in Medicine
(a) (b) (c)
Figure 4 Registration between (a) SPECT and (b) MR to give the final overlaid image in (c)
(a) (b) (c)
Figure 5 The ST regions in (a) the original MR T1-weighted image and the corresponding segmentation results by (b) manual delineationand (c) the GVF snake
Table 2 Correspondence betweenmanual delineation and theGVFsnake result for each observer
JI () Rater A-GVFsnake
Rater B-GVFsnake
Rater C-GVFsnake
Case 1 644 plusmn 90 562 plusmn 55 653 plusmn 48
Case 2 686 plusmn 16 577 plusmn 51 654 plusmn 65
Case 3 611 plusmn 38 515 plusmn 31 599 plusmn 37
defined by the same specialist in the GVF snake processTable 2 shows the correspondence with GVF segmentationWe used the paired 119905-test to evaluate the significance levelbetween the JIs derived from the interrater comparison andthose from the rater-GVF comparison for each slice in eachcase The insignificant differences (119875 = 0124 under a 95confidence interval) suggest that the segmentation accuracyusing the GVF snake is compatible with the level of manualdrawing
Rater A the chief neurologist was required to conductthe manual drawing twice The JI values of the two delin-eations are listed in the middle column of Table 3 showing
Table 3 Correspondence between two repeated conductions ofeach method
that the correspondence declined along with the slice num-ber Instead of segmentation solely by hand the GVF snakewas also applied twice to the same set of images Only thefirst slice required an initial contour manually defined by therater each time The JI values of repeated conduction of theGVF snake are also listed in the third column of Table 3suggesting more stable results than those from slice-by-slicemanual drawing
Computational and Mathematical Methods in Medicine 7
In comparison with the index of overlap (similar to JI)between hand-drawing and computer-aided segmentationreported in the literature [9ndash11] the JI values obtained inour study were relatively low This could be due to the extraregion that is the putamen involved in our study Thesegmentation target of the previous reports is focused onthe caudate nucleuses that are next to ventricles with greatercontrast (Figure 1) and therefore more easily identified Incomparison with the caudate nucleuses the low contrastof the putamen to the surrounding tissue increases thedifficulty of extraction Using expert hand drawing as thecomparison basis seems to be the only choice in currentstudies since there is no gold standard to determine theabsolute accuracy of segmentation of the ST regions dueto individual-dependent variation in the brain structureHowever we demonstrated that significant interobserver andintraobserver variability in such a decision exists even amongthe well-trained neurologists participating in our studywhich was overlooked in previous studies The inconsistencyin decision-making could be incurred by the small size (inthe order of 100 pixels) of the structures as compared with theimaging resolution and image noise Under a compatible levelof precision as shown in Tables 1 and 2 we demonstrated thatthe reproducibility and consistency improved when using theGVF snake segmentationmethod In addition to stability theGVF snake can save labor and provide a more efficient waythan previous studies to define the ST regions contours inconsecutive slices as it only requires an initial contour drawnby hand in the first slice
33 Binding Ratio Calculation In the final stage of evaluationof the reliability of the proposedmethod after completing theMR and SPECT image registration the BRs were also derivedfrom the segmented ST regions in the SPECT images usingthe proposed method to compare with those obtained usingcommercial software (Siemens Medical Systems KnoxvilleTN USA) in which the ST regions were manually outlinedby an expert The BR was calculated by normalizing themean intensity in the ST regions by the mean intensity inthe occipital cortices Linear regression analysis (Figure 6)revealed a close correlation (CC = 0874 under 95confidence interval) between the BRs derived by the twosystems
4 Conclusions
To calculate the regional TRODAT-1 binding ratio in SPECTstudies accurate and repeatable extraction of the ST regionsfrom MR images is required to indirectly define the cor-responding regions in the SPECT images Segmentationdirectly on the SPECT image is not applicable in this casebecause it distorts the ST regions Clinical routine tends toapply manual delineation of the ST regions which is proneto errors incurred through interobserver and intraobservervariability Previous researchers have developed several seg-mentation algorithms to complete similar tasks where expertdecisions for anatomical and morphological informationwere still necessary to refine the results As the localization
1
11
12
13
14
15
16
1 12 14 16 18 2 22 24 26
Our
met
hod
Commercial software
y = 0466x + 0415
R2= 0764
Correlation coefficient = 0874
Figure 6 Linear regression analysis between the BRs derived by ourmethod and those by the commercial software
of the ST regions is a knowledge-driven task the proposedmethod allowed the expert to assign the initial contourin the proper location and applied the gradient-vector-flow snake to approach the real contours In such a waythe complexity of the algorithm can be reduced and theefficiency of segmentation can be increased Results fromthree normal subjects showed a higher reproducibility ofthe proposed method than manual segmentation undercompatible segmentation accuracy The MR images withsegmented ST regions were overlaid on the SPECT imagesusing our previously developed registration algorithm tocalculate the TRODAT-1 BRThe derived BRs correlated wellwith those derived using commercial software suggesting agood reliability of the proposed method
Acknowledgments
The authors would like to thank Dr Chen Nai-Ching andDr Chi-Wei Huang from the Department of NeurologyChang Gung Memorial Hospital Kaohsiung Medical CenterKaohsiung Taiwan for their kind assistance in the task ofmanual drawing of the ST regions in this study
References
[1] S Asenbaum ldquoNuclear medicine in neurodegenerative disor-dersrdquo Imaging Decisions MRI vol 6 pp 19ndash28 2002
[2] K L Chou H I Hurtig M B Stern et al ldquoDiagnostic accuracyof [ 99mTc]TRODAT-1 SPECT imaging in early Parkinsonrsquosdiseaserdquo Parkinsonism and Related Disorders vol 10 no 6 pp375ndash379 2004
[3] J L Cummings C Henchcliffe S Schaier T Simuni AWaxman and P Kemp ldquoThe role of dopaminergic imaging inpatients with symptoms of dopaminergic system neurodegen-erationrdquo Brain vol 134 no 11 pp 3146ndash3166 2011
[4] A Siderowf A Newberg K L Chou et al ldquo[ 99mTc]TRODAT-1 SPECT imaging correlates with odor identification in earlyParkinson diseaserdquo Neurology vol 64 no 10 pp 1716ndash17202005
8 Computational and Mathematical Methods in Medicine
[5] Y-H Weng T-C Yen M-C Chen et al ldquoSensitivity andspecificity of 99mTc-TRODAT-1 SPECT imaging in differenti-ating patients with idiopathic Parkinsonrsquos disease from healthysubjectsrdquo Journal of NuclearMedicine vol 45 no 3 pp 393ndash4012004
[6] W-S Huang S-Z Lin J-C Lin S-P Wey G Ting and R-SLiu ldquoEvaluation of early-stage Parkinsonrsquos disease with 99mTc-TRODAT-1 imagingrdquo Journal of NuclearMedicine vol 42 no 9pp 1303ndash1308 2001
[7] P D Acton P T Meyer P D Mozley K Plossl and H F KungldquoSimplified quantification of dopamine transporters in humansusing [ 99mTc]TRODAT-1 and single-photon emission tomog-raphyrdquo European Journal of Nuclear Medicine and MolecularImaging vol 27 no 11 pp 1714ndash1718 2000
[8] P D Acton P D Mozley and H F Kung ldquoLogistic discrimi-nant parametric mapping a novel method for the pixel-baseddifferential diagnosis of Parkinsonrsquos diseaserdquo European Journalof Nuclear Medicine and Molecular Imaging vol 26 no 11 pp1413ndash1423 1999
[9] A J Worth N Makris M R Patti et al ldquoPrecise segmentationof the lateral ventricles and caudate nucleus in mr brain imagesusing anatomically driven histogramsrdquo IEEE Transactions onMedical Imaging vol 17 no 2 pp 303ndash310 1998
[10] V Barra and J-Y Boire ldquoAutomatic segmentation of subcorticalbrain structures in MR images using information fusionrdquo IEEETransactions on Medical Imaging vol 20 no 7 pp 549ndash5582001
[11] Y Xia K Bettinger L Shen and A L Reiss ldquoAutomaticsegmentation of the caudate nucleus from human brain MRimagesrdquo IEEE Transactions on Medical Imaging vol 26 no 4pp 509ndash517 2007
[12] C-F Jiang C-C Chang and S-H Huang ldquoRegions of interestextraction from SPECT images for neural degeneration assess-ment using multimodality image fusionrdquo MultidimensionalSystems and Signal Processing vol 23 pp 437ndash449 2012
[13] A P Dhawan ldquoImage registrationrdquo in Medical Image AnalysisM Akay Ed pp 251ndash276 IEEE Press Piscataway NJ USA2003
[14] A P Dhawan L K Arata A V Levy and J Mantil ldquoIterativeprincipal axes registrationmethod for analysis ofMR-PETbrainimagesrdquo IEEE Transactions on Biomedical Engineering vol 42no 11 pp 1079ndash1087 1995
[15] C-F Jiang C-H Huang and S-T Yang ldquoUsingmaximal cross-section detection for the registration of 3D image data of theheadrdquo Journal of Medical and Biological Engineering vol 31 no3 pp 217ndash226 2011
[16] MKass AWitkin andD Terzopoulos ldquoSnakes active contourmodelsrdquo International Journal of Computer Vision vol 1 no 4pp 321ndash331 1988
[17] L D Cohen and I Cohen ldquoFinite-element methods for activecontour models and balloons for 2-D and 3-D imagesrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol15 no 11 pp 1131ndash1147 1993
[18] B Leroy I Herlin and L Cohen ldquoMulti-resolution algorithmsfor active contour modelsrdquo in Proceedings of the 12th Inter-national Conference on Analysis and Optimization of SystemsImages (ICAOS rsquo96) pp 58ndash65 1996
[19] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998
[20] B Erkol R H Moss R J Stanley W V Stoecker and EHvatum ldquoAutomatic lesion boundary detection in dermoscopy
images using gradient vector flow snakesrdquo Skin Research andTechnology vol 11 no 1 pp 17ndash26 2005
[21] J Tang and S T Acton ldquoVessel boundary tracking for intravitalmicroscopy via multiscale gradient vector flow snakesrdquo IEEETransactions on Biomedical Engineering vol 51 no 2 pp 316ndash324 2004
[22] J Tang S Millington S T Acton J Crandall and S HurwitzldquoSurface extraction and thickness measurement of the articularcartilage fromMR images using directional gradient vector flowsnakesrdquo IEEE Transactions on Biomedical Engineering vol 53no 5 pp 896ndash907 2006
[23] N Tanki K Murase M Kumashiro et al ldquoQuantification ofleft ventricular volumes from cardiac cine MRI using activecontour model combined with gradient vector flowrdquo MagneticResonance in Medical Sciences vol 4 no 4 pp 191ndash196 2005
[24] C Xu and J L Prince ldquoGradient vector flow deformable mod-elsrdquo in Handbook of Medical Imaging pp 159ndash169 AcademicPress 2000
Computational and Mathematical Methods in Medicine 5
the internal forces Even though several dynamic externalforces have been proposed to alleviate such a limitation ofthe static external forces they also raised other problemsincreasing the calculation complexity or leading to uncon-trollable deformation [17 18] A new static external forcecalled gradient-vector flow (GVF) adding the directionalproperty into the original image gradient map was proposedby Xu and Prince to improve the performance of the staticsnake in concave edge detection [19] Several reports havedemonstrated the success of applying the GVF snake tomedical image segmentation [20ndash23] including brain MRI[24] This encouraged us to apply the GVF to segment the STregions in our study
The gradient-vector-flow field is defined as k(119909 119910) =[119906(119909 119910) V(119909 119910)] such that the external energy functionbecomes
where 120583 is a parameter to control the degree of smoothness ofthe gradient-vector-flow field and nabla119891 is an edge map derivedfrom the original image 119891(119909 119910)
To solve the equation numerically by discretization anditeration let 119899 indicate the times of iteration and theincrements in 119909 119910 and 119905 are all equal to 1 The relation ofvector flows from the current to the next position can bederived as
There are 4 parameters determined empirically to obtainthe optimal results in the approach The elasticity parameter(120572) and the rigidity parameter (120573) in (2) were set to be 01 and02 respectively The parameter (120574) in (6) was set to be 1 Theexternal force weight (120583) in (7) was set to be 05
3 Results and Discussion
31 Registration Results An example is given in Figure 3 toillustrate the use of our developed interface to detect the slicescontaining the MCSA from the CT and MR volumes andregister these two images through the GHT Once the rigidtransform with the registration parameters had been appliedto the MR image volume it can directly match the SPECTvolume as shown in Figure 4 The registration accuracyreached 9648 Our previous study quantitatively evaluated
Figure 3 An example of registration of the CT image (middle smallpanel) and theMR image (right small panel) to render the final fusedimage (left large panel)The red and blue lines in the small panels arethe detected boundaries and the contour of the head
Table 1 Correspondence of the manual delineation between theobservers
JI () Rater A-Rater B Rater B-Rater C Rater A-Rater CCase 1 604 plusmn 48 621 plusmn 11 756 plusmn 21
Case 2 543 plusmn 45 547 plusmn 56 718 plusmn 52
Case 3 580 plusmn 47 581 plusmn 54 766 plusmn 52
the registration accuracy of the proposed method better thanthe results obtained solely using the PAR method or directlyregistering SPECT with MR images [12]
32 Segmentation Results The expert delineation and theGVF segmentation of the ST regions containing two pairsof the caudate nucleus and putamen are given in Figure 5 Aquantitative comparison of these twomethods is given below
We used the Jaccard index (JI) to quantify the degree ofmatch between the two corresponding ROIsThe JI is definedas the ratio of the intersection of two volumes Ω
1and Ω
2by
the union of them If the two volumes completely overlap theJI value is equal to 100
JI =1003816100381610038161003816Ω1cap Ω2
1003816100381610038161003816
1003816100381610038161003816Ω1cup Ω2
1003816100381610038161003816
times 100 (9)
The five sequential axial slices containing the ST fromthree normal cases were recruited in the comparative eval-uation Three neurologists first manually delineated the STregions including the caudate nucleus and putamen on twolateral sides of the brain The intrarater correspondences interms of the mean and standard deviation of the JI valuesfrom the five slices are listed in Table 1 suggesting greatdifferences between observers It was found that Raters A andChad the highest correspondence with a JI value greater than70
The JI values were also derived by mapping the manuallydefined contours by each rater into the GVF segmentedresults The initial contour of the first slice in each case was
6 Computational and Mathematical Methods in Medicine
(a) (b) (c)
Figure 4 Registration between (a) SPECT and (b) MR to give the final overlaid image in (c)
(a) (b) (c)
Figure 5 The ST regions in (a) the original MR T1-weighted image and the corresponding segmentation results by (b) manual delineationand (c) the GVF snake
Table 2 Correspondence betweenmanual delineation and theGVFsnake result for each observer
JI () Rater A-GVFsnake
Rater B-GVFsnake
Rater C-GVFsnake
Case 1 644 plusmn 90 562 plusmn 55 653 plusmn 48
Case 2 686 plusmn 16 577 plusmn 51 654 plusmn 65
Case 3 611 plusmn 38 515 plusmn 31 599 plusmn 37
defined by the same specialist in the GVF snake processTable 2 shows the correspondence with GVF segmentationWe used the paired 119905-test to evaluate the significance levelbetween the JIs derived from the interrater comparison andthose from the rater-GVF comparison for each slice in eachcase The insignificant differences (119875 = 0124 under a 95confidence interval) suggest that the segmentation accuracyusing the GVF snake is compatible with the level of manualdrawing
Rater A the chief neurologist was required to conductthe manual drawing twice The JI values of the two delin-eations are listed in the middle column of Table 3 showing
Table 3 Correspondence between two repeated conductions ofeach method
that the correspondence declined along with the slice num-ber Instead of segmentation solely by hand the GVF snakewas also applied twice to the same set of images Only thefirst slice required an initial contour manually defined by therater each time The JI values of repeated conduction of theGVF snake are also listed in the third column of Table 3suggesting more stable results than those from slice-by-slicemanual drawing
Computational and Mathematical Methods in Medicine 7
In comparison with the index of overlap (similar to JI)between hand-drawing and computer-aided segmentationreported in the literature [9ndash11] the JI values obtained inour study were relatively low This could be due to the extraregion that is the putamen involved in our study Thesegmentation target of the previous reports is focused onthe caudate nucleuses that are next to ventricles with greatercontrast (Figure 1) and therefore more easily identified Incomparison with the caudate nucleuses the low contrastof the putamen to the surrounding tissue increases thedifficulty of extraction Using expert hand drawing as thecomparison basis seems to be the only choice in currentstudies since there is no gold standard to determine theabsolute accuracy of segmentation of the ST regions dueto individual-dependent variation in the brain structureHowever we demonstrated that significant interobserver andintraobserver variability in such a decision exists even amongthe well-trained neurologists participating in our studywhich was overlooked in previous studies The inconsistencyin decision-making could be incurred by the small size (inthe order of 100 pixels) of the structures as compared with theimaging resolution and image noise Under a compatible levelof precision as shown in Tables 1 and 2 we demonstrated thatthe reproducibility and consistency improved when using theGVF snake segmentationmethod In addition to stability theGVF snake can save labor and provide a more efficient waythan previous studies to define the ST regions contours inconsecutive slices as it only requires an initial contour drawnby hand in the first slice
33 Binding Ratio Calculation In the final stage of evaluationof the reliability of the proposedmethod after completing theMR and SPECT image registration the BRs were also derivedfrom the segmented ST regions in the SPECT images usingthe proposed method to compare with those obtained usingcommercial software (Siemens Medical Systems KnoxvilleTN USA) in which the ST regions were manually outlinedby an expert The BR was calculated by normalizing themean intensity in the ST regions by the mean intensity inthe occipital cortices Linear regression analysis (Figure 6)revealed a close correlation (CC = 0874 under 95confidence interval) between the BRs derived by the twosystems
4 Conclusions
To calculate the regional TRODAT-1 binding ratio in SPECTstudies accurate and repeatable extraction of the ST regionsfrom MR images is required to indirectly define the cor-responding regions in the SPECT images Segmentationdirectly on the SPECT image is not applicable in this casebecause it distorts the ST regions Clinical routine tends toapply manual delineation of the ST regions which is proneto errors incurred through interobserver and intraobservervariability Previous researchers have developed several seg-mentation algorithms to complete similar tasks where expertdecisions for anatomical and morphological informationwere still necessary to refine the results As the localization
1
11
12
13
14
15
16
1 12 14 16 18 2 22 24 26
Our
met
hod
Commercial software
y = 0466x + 0415
R2= 0764
Correlation coefficient = 0874
Figure 6 Linear regression analysis between the BRs derived by ourmethod and those by the commercial software
of the ST regions is a knowledge-driven task the proposedmethod allowed the expert to assign the initial contourin the proper location and applied the gradient-vector-flow snake to approach the real contours In such a waythe complexity of the algorithm can be reduced and theefficiency of segmentation can be increased Results fromthree normal subjects showed a higher reproducibility ofthe proposed method than manual segmentation undercompatible segmentation accuracy The MR images withsegmented ST regions were overlaid on the SPECT imagesusing our previously developed registration algorithm tocalculate the TRODAT-1 BRThe derived BRs correlated wellwith those derived using commercial software suggesting agood reliability of the proposed method
Acknowledgments
The authors would like to thank Dr Chen Nai-Ching andDr Chi-Wei Huang from the Department of NeurologyChang Gung Memorial Hospital Kaohsiung Medical CenterKaohsiung Taiwan for their kind assistance in the task ofmanual drawing of the ST regions in this study
References
[1] S Asenbaum ldquoNuclear medicine in neurodegenerative disor-dersrdquo Imaging Decisions MRI vol 6 pp 19ndash28 2002
[2] K L Chou H I Hurtig M B Stern et al ldquoDiagnostic accuracyof [ 99mTc]TRODAT-1 SPECT imaging in early Parkinsonrsquosdiseaserdquo Parkinsonism and Related Disorders vol 10 no 6 pp375ndash379 2004
[3] J L Cummings C Henchcliffe S Schaier T Simuni AWaxman and P Kemp ldquoThe role of dopaminergic imaging inpatients with symptoms of dopaminergic system neurodegen-erationrdquo Brain vol 134 no 11 pp 3146ndash3166 2011
[4] A Siderowf A Newberg K L Chou et al ldquo[ 99mTc]TRODAT-1 SPECT imaging correlates with odor identification in earlyParkinson diseaserdquo Neurology vol 64 no 10 pp 1716ndash17202005
8 Computational and Mathematical Methods in Medicine
[5] Y-H Weng T-C Yen M-C Chen et al ldquoSensitivity andspecificity of 99mTc-TRODAT-1 SPECT imaging in differenti-ating patients with idiopathic Parkinsonrsquos disease from healthysubjectsrdquo Journal of NuclearMedicine vol 45 no 3 pp 393ndash4012004
[6] W-S Huang S-Z Lin J-C Lin S-P Wey G Ting and R-SLiu ldquoEvaluation of early-stage Parkinsonrsquos disease with 99mTc-TRODAT-1 imagingrdquo Journal of NuclearMedicine vol 42 no 9pp 1303ndash1308 2001
[7] P D Acton P T Meyer P D Mozley K Plossl and H F KungldquoSimplified quantification of dopamine transporters in humansusing [ 99mTc]TRODAT-1 and single-photon emission tomog-raphyrdquo European Journal of Nuclear Medicine and MolecularImaging vol 27 no 11 pp 1714ndash1718 2000
[8] P D Acton P D Mozley and H F Kung ldquoLogistic discrimi-nant parametric mapping a novel method for the pixel-baseddifferential diagnosis of Parkinsonrsquos diseaserdquo European Journalof Nuclear Medicine and Molecular Imaging vol 26 no 11 pp1413ndash1423 1999
[9] A J Worth N Makris M R Patti et al ldquoPrecise segmentationof the lateral ventricles and caudate nucleus in mr brain imagesusing anatomically driven histogramsrdquo IEEE Transactions onMedical Imaging vol 17 no 2 pp 303ndash310 1998
[10] V Barra and J-Y Boire ldquoAutomatic segmentation of subcorticalbrain structures in MR images using information fusionrdquo IEEETransactions on Medical Imaging vol 20 no 7 pp 549ndash5582001
[11] Y Xia K Bettinger L Shen and A L Reiss ldquoAutomaticsegmentation of the caudate nucleus from human brain MRimagesrdquo IEEE Transactions on Medical Imaging vol 26 no 4pp 509ndash517 2007
[12] C-F Jiang C-C Chang and S-H Huang ldquoRegions of interestextraction from SPECT images for neural degeneration assess-ment using multimodality image fusionrdquo MultidimensionalSystems and Signal Processing vol 23 pp 437ndash449 2012
[13] A P Dhawan ldquoImage registrationrdquo in Medical Image AnalysisM Akay Ed pp 251ndash276 IEEE Press Piscataway NJ USA2003
[14] A P Dhawan L K Arata A V Levy and J Mantil ldquoIterativeprincipal axes registrationmethod for analysis ofMR-PETbrainimagesrdquo IEEE Transactions on Biomedical Engineering vol 42no 11 pp 1079ndash1087 1995
[15] C-F Jiang C-H Huang and S-T Yang ldquoUsingmaximal cross-section detection for the registration of 3D image data of theheadrdquo Journal of Medical and Biological Engineering vol 31 no3 pp 217ndash226 2011
[16] MKass AWitkin andD Terzopoulos ldquoSnakes active contourmodelsrdquo International Journal of Computer Vision vol 1 no 4pp 321ndash331 1988
[17] L D Cohen and I Cohen ldquoFinite-element methods for activecontour models and balloons for 2-D and 3-D imagesrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol15 no 11 pp 1131ndash1147 1993
[18] B Leroy I Herlin and L Cohen ldquoMulti-resolution algorithmsfor active contour modelsrdquo in Proceedings of the 12th Inter-national Conference on Analysis and Optimization of SystemsImages (ICAOS rsquo96) pp 58ndash65 1996
[19] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998
[20] B Erkol R H Moss R J Stanley W V Stoecker and EHvatum ldquoAutomatic lesion boundary detection in dermoscopy
images using gradient vector flow snakesrdquo Skin Research andTechnology vol 11 no 1 pp 17ndash26 2005
[21] J Tang and S T Acton ldquoVessel boundary tracking for intravitalmicroscopy via multiscale gradient vector flow snakesrdquo IEEETransactions on Biomedical Engineering vol 51 no 2 pp 316ndash324 2004
[22] J Tang S Millington S T Acton J Crandall and S HurwitzldquoSurface extraction and thickness measurement of the articularcartilage fromMR images using directional gradient vector flowsnakesrdquo IEEE Transactions on Biomedical Engineering vol 53no 5 pp 896ndash907 2006
[23] N Tanki K Murase M Kumashiro et al ldquoQuantification ofleft ventricular volumes from cardiac cine MRI using activecontour model combined with gradient vector flowrdquo MagneticResonance in Medical Sciences vol 4 no 4 pp 191ndash196 2005
[24] C Xu and J L Prince ldquoGradient vector flow deformable mod-elsrdquo in Handbook of Medical Imaging pp 159ndash169 AcademicPress 2000
6 Computational and Mathematical Methods in Medicine
(a) (b) (c)
Figure 4 Registration between (a) SPECT and (b) MR to give the final overlaid image in (c)
(a) (b) (c)
Figure 5 The ST regions in (a) the original MR T1-weighted image and the corresponding segmentation results by (b) manual delineationand (c) the GVF snake
Table 2 Correspondence betweenmanual delineation and theGVFsnake result for each observer
JI () Rater A-GVFsnake
Rater B-GVFsnake
Rater C-GVFsnake
Case 1 644 plusmn 90 562 plusmn 55 653 plusmn 48
Case 2 686 plusmn 16 577 plusmn 51 654 plusmn 65
Case 3 611 plusmn 38 515 plusmn 31 599 plusmn 37
defined by the same specialist in the GVF snake processTable 2 shows the correspondence with GVF segmentationWe used the paired 119905-test to evaluate the significance levelbetween the JIs derived from the interrater comparison andthose from the rater-GVF comparison for each slice in eachcase The insignificant differences (119875 = 0124 under a 95confidence interval) suggest that the segmentation accuracyusing the GVF snake is compatible with the level of manualdrawing
Rater A the chief neurologist was required to conductthe manual drawing twice The JI values of the two delin-eations are listed in the middle column of Table 3 showing
Table 3 Correspondence between two repeated conductions ofeach method
that the correspondence declined along with the slice num-ber Instead of segmentation solely by hand the GVF snakewas also applied twice to the same set of images Only thefirst slice required an initial contour manually defined by therater each time The JI values of repeated conduction of theGVF snake are also listed in the third column of Table 3suggesting more stable results than those from slice-by-slicemanual drawing
Computational and Mathematical Methods in Medicine 7
In comparison with the index of overlap (similar to JI)between hand-drawing and computer-aided segmentationreported in the literature [9ndash11] the JI values obtained inour study were relatively low This could be due to the extraregion that is the putamen involved in our study Thesegmentation target of the previous reports is focused onthe caudate nucleuses that are next to ventricles with greatercontrast (Figure 1) and therefore more easily identified Incomparison with the caudate nucleuses the low contrastof the putamen to the surrounding tissue increases thedifficulty of extraction Using expert hand drawing as thecomparison basis seems to be the only choice in currentstudies since there is no gold standard to determine theabsolute accuracy of segmentation of the ST regions dueto individual-dependent variation in the brain structureHowever we demonstrated that significant interobserver andintraobserver variability in such a decision exists even amongthe well-trained neurologists participating in our studywhich was overlooked in previous studies The inconsistencyin decision-making could be incurred by the small size (inthe order of 100 pixels) of the structures as compared with theimaging resolution and image noise Under a compatible levelof precision as shown in Tables 1 and 2 we demonstrated thatthe reproducibility and consistency improved when using theGVF snake segmentationmethod In addition to stability theGVF snake can save labor and provide a more efficient waythan previous studies to define the ST regions contours inconsecutive slices as it only requires an initial contour drawnby hand in the first slice
33 Binding Ratio Calculation In the final stage of evaluationof the reliability of the proposedmethod after completing theMR and SPECT image registration the BRs were also derivedfrom the segmented ST regions in the SPECT images usingthe proposed method to compare with those obtained usingcommercial software (Siemens Medical Systems KnoxvilleTN USA) in which the ST regions were manually outlinedby an expert The BR was calculated by normalizing themean intensity in the ST regions by the mean intensity inthe occipital cortices Linear regression analysis (Figure 6)revealed a close correlation (CC = 0874 under 95confidence interval) between the BRs derived by the twosystems
4 Conclusions
To calculate the regional TRODAT-1 binding ratio in SPECTstudies accurate and repeatable extraction of the ST regionsfrom MR images is required to indirectly define the cor-responding regions in the SPECT images Segmentationdirectly on the SPECT image is not applicable in this casebecause it distorts the ST regions Clinical routine tends toapply manual delineation of the ST regions which is proneto errors incurred through interobserver and intraobservervariability Previous researchers have developed several seg-mentation algorithms to complete similar tasks where expertdecisions for anatomical and morphological informationwere still necessary to refine the results As the localization
1
11
12
13
14
15
16
1 12 14 16 18 2 22 24 26
Our
met
hod
Commercial software
y = 0466x + 0415
R2= 0764
Correlation coefficient = 0874
Figure 6 Linear regression analysis between the BRs derived by ourmethod and those by the commercial software
of the ST regions is a knowledge-driven task the proposedmethod allowed the expert to assign the initial contourin the proper location and applied the gradient-vector-flow snake to approach the real contours In such a waythe complexity of the algorithm can be reduced and theefficiency of segmentation can be increased Results fromthree normal subjects showed a higher reproducibility ofthe proposed method than manual segmentation undercompatible segmentation accuracy The MR images withsegmented ST regions were overlaid on the SPECT imagesusing our previously developed registration algorithm tocalculate the TRODAT-1 BRThe derived BRs correlated wellwith those derived using commercial software suggesting agood reliability of the proposed method
Acknowledgments
The authors would like to thank Dr Chen Nai-Ching andDr Chi-Wei Huang from the Department of NeurologyChang Gung Memorial Hospital Kaohsiung Medical CenterKaohsiung Taiwan for their kind assistance in the task ofmanual drawing of the ST regions in this study
References
[1] S Asenbaum ldquoNuclear medicine in neurodegenerative disor-dersrdquo Imaging Decisions MRI vol 6 pp 19ndash28 2002
[2] K L Chou H I Hurtig M B Stern et al ldquoDiagnostic accuracyof [ 99mTc]TRODAT-1 SPECT imaging in early Parkinsonrsquosdiseaserdquo Parkinsonism and Related Disorders vol 10 no 6 pp375ndash379 2004
[3] J L Cummings C Henchcliffe S Schaier T Simuni AWaxman and P Kemp ldquoThe role of dopaminergic imaging inpatients with symptoms of dopaminergic system neurodegen-erationrdquo Brain vol 134 no 11 pp 3146ndash3166 2011
[4] A Siderowf A Newberg K L Chou et al ldquo[ 99mTc]TRODAT-1 SPECT imaging correlates with odor identification in earlyParkinson diseaserdquo Neurology vol 64 no 10 pp 1716ndash17202005
8 Computational and Mathematical Methods in Medicine
[5] Y-H Weng T-C Yen M-C Chen et al ldquoSensitivity andspecificity of 99mTc-TRODAT-1 SPECT imaging in differenti-ating patients with idiopathic Parkinsonrsquos disease from healthysubjectsrdquo Journal of NuclearMedicine vol 45 no 3 pp 393ndash4012004
[6] W-S Huang S-Z Lin J-C Lin S-P Wey G Ting and R-SLiu ldquoEvaluation of early-stage Parkinsonrsquos disease with 99mTc-TRODAT-1 imagingrdquo Journal of NuclearMedicine vol 42 no 9pp 1303ndash1308 2001
[7] P D Acton P T Meyer P D Mozley K Plossl and H F KungldquoSimplified quantification of dopamine transporters in humansusing [ 99mTc]TRODAT-1 and single-photon emission tomog-raphyrdquo European Journal of Nuclear Medicine and MolecularImaging vol 27 no 11 pp 1714ndash1718 2000
[8] P D Acton P D Mozley and H F Kung ldquoLogistic discrimi-nant parametric mapping a novel method for the pixel-baseddifferential diagnosis of Parkinsonrsquos diseaserdquo European Journalof Nuclear Medicine and Molecular Imaging vol 26 no 11 pp1413ndash1423 1999
[9] A J Worth N Makris M R Patti et al ldquoPrecise segmentationof the lateral ventricles and caudate nucleus in mr brain imagesusing anatomically driven histogramsrdquo IEEE Transactions onMedical Imaging vol 17 no 2 pp 303ndash310 1998
[10] V Barra and J-Y Boire ldquoAutomatic segmentation of subcorticalbrain structures in MR images using information fusionrdquo IEEETransactions on Medical Imaging vol 20 no 7 pp 549ndash5582001
[11] Y Xia K Bettinger L Shen and A L Reiss ldquoAutomaticsegmentation of the caudate nucleus from human brain MRimagesrdquo IEEE Transactions on Medical Imaging vol 26 no 4pp 509ndash517 2007
[12] C-F Jiang C-C Chang and S-H Huang ldquoRegions of interestextraction from SPECT images for neural degeneration assess-ment using multimodality image fusionrdquo MultidimensionalSystems and Signal Processing vol 23 pp 437ndash449 2012
[13] A P Dhawan ldquoImage registrationrdquo in Medical Image AnalysisM Akay Ed pp 251ndash276 IEEE Press Piscataway NJ USA2003
[14] A P Dhawan L K Arata A V Levy and J Mantil ldquoIterativeprincipal axes registrationmethod for analysis ofMR-PETbrainimagesrdquo IEEE Transactions on Biomedical Engineering vol 42no 11 pp 1079ndash1087 1995
[15] C-F Jiang C-H Huang and S-T Yang ldquoUsingmaximal cross-section detection for the registration of 3D image data of theheadrdquo Journal of Medical and Biological Engineering vol 31 no3 pp 217ndash226 2011
[16] MKass AWitkin andD Terzopoulos ldquoSnakes active contourmodelsrdquo International Journal of Computer Vision vol 1 no 4pp 321ndash331 1988
[17] L D Cohen and I Cohen ldquoFinite-element methods for activecontour models and balloons for 2-D and 3-D imagesrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol15 no 11 pp 1131ndash1147 1993
[18] B Leroy I Herlin and L Cohen ldquoMulti-resolution algorithmsfor active contour modelsrdquo in Proceedings of the 12th Inter-national Conference on Analysis and Optimization of SystemsImages (ICAOS rsquo96) pp 58ndash65 1996
[19] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998
[20] B Erkol R H Moss R J Stanley W V Stoecker and EHvatum ldquoAutomatic lesion boundary detection in dermoscopy
images using gradient vector flow snakesrdquo Skin Research andTechnology vol 11 no 1 pp 17ndash26 2005
[21] J Tang and S T Acton ldquoVessel boundary tracking for intravitalmicroscopy via multiscale gradient vector flow snakesrdquo IEEETransactions on Biomedical Engineering vol 51 no 2 pp 316ndash324 2004
[22] J Tang S Millington S T Acton J Crandall and S HurwitzldquoSurface extraction and thickness measurement of the articularcartilage fromMR images using directional gradient vector flowsnakesrdquo IEEE Transactions on Biomedical Engineering vol 53no 5 pp 896ndash907 2006
[23] N Tanki K Murase M Kumashiro et al ldquoQuantification ofleft ventricular volumes from cardiac cine MRI using activecontour model combined with gradient vector flowrdquo MagneticResonance in Medical Sciences vol 4 no 4 pp 191ndash196 2005
[24] C Xu and J L Prince ldquoGradient vector flow deformable mod-elsrdquo in Handbook of Medical Imaging pp 159ndash169 AcademicPress 2000
Computational and Mathematical Methods in Medicine 7
In comparison with the index of overlap (similar to JI)between hand-drawing and computer-aided segmentationreported in the literature [9ndash11] the JI values obtained inour study were relatively low This could be due to the extraregion that is the putamen involved in our study Thesegmentation target of the previous reports is focused onthe caudate nucleuses that are next to ventricles with greatercontrast (Figure 1) and therefore more easily identified Incomparison with the caudate nucleuses the low contrastof the putamen to the surrounding tissue increases thedifficulty of extraction Using expert hand drawing as thecomparison basis seems to be the only choice in currentstudies since there is no gold standard to determine theabsolute accuracy of segmentation of the ST regions dueto individual-dependent variation in the brain structureHowever we demonstrated that significant interobserver andintraobserver variability in such a decision exists even amongthe well-trained neurologists participating in our studywhich was overlooked in previous studies The inconsistencyin decision-making could be incurred by the small size (inthe order of 100 pixels) of the structures as compared with theimaging resolution and image noise Under a compatible levelof precision as shown in Tables 1 and 2 we demonstrated thatthe reproducibility and consistency improved when using theGVF snake segmentationmethod In addition to stability theGVF snake can save labor and provide a more efficient waythan previous studies to define the ST regions contours inconsecutive slices as it only requires an initial contour drawnby hand in the first slice
33 Binding Ratio Calculation In the final stage of evaluationof the reliability of the proposedmethod after completing theMR and SPECT image registration the BRs were also derivedfrom the segmented ST regions in the SPECT images usingthe proposed method to compare with those obtained usingcommercial software (Siemens Medical Systems KnoxvilleTN USA) in which the ST regions were manually outlinedby an expert The BR was calculated by normalizing themean intensity in the ST regions by the mean intensity inthe occipital cortices Linear regression analysis (Figure 6)revealed a close correlation (CC = 0874 under 95confidence interval) between the BRs derived by the twosystems
4 Conclusions
To calculate the regional TRODAT-1 binding ratio in SPECTstudies accurate and repeatable extraction of the ST regionsfrom MR images is required to indirectly define the cor-responding regions in the SPECT images Segmentationdirectly on the SPECT image is not applicable in this casebecause it distorts the ST regions Clinical routine tends toapply manual delineation of the ST regions which is proneto errors incurred through interobserver and intraobservervariability Previous researchers have developed several seg-mentation algorithms to complete similar tasks where expertdecisions for anatomical and morphological informationwere still necessary to refine the results As the localization
1
11
12
13
14
15
16
1 12 14 16 18 2 22 24 26
Our
met
hod
Commercial software
y = 0466x + 0415
R2= 0764
Correlation coefficient = 0874
Figure 6 Linear regression analysis between the BRs derived by ourmethod and those by the commercial software
of the ST regions is a knowledge-driven task the proposedmethod allowed the expert to assign the initial contourin the proper location and applied the gradient-vector-flow snake to approach the real contours In such a waythe complexity of the algorithm can be reduced and theefficiency of segmentation can be increased Results fromthree normal subjects showed a higher reproducibility ofthe proposed method than manual segmentation undercompatible segmentation accuracy The MR images withsegmented ST regions were overlaid on the SPECT imagesusing our previously developed registration algorithm tocalculate the TRODAT-1 BRThe derived BRs correlated wellwith those derived using commercial software suggesting agood reliability of the proposed method
Acknowledgments
The authors would like to thank Dr Chen Nai-Ching andDr Chi-Wei Huang from the Department of NeurologyChang Gung Memorial Hospital Kaohsiung Medical CenterKaohsiung Taiwan for their kind assistance in the task ofmanual drawing of the ST regions in this study
References
[1] S Asenbaum ldquoNuclear medicine in neurodegenerative disor-dersrdquo Imaging Decisions MRI vol 6 pp 19ndash28 2002
[2] K L Chou H I Hurtig M B Stern et al ldquoDiagnostic accuracyof [ 99mTc]TRODAT-1 SPECT imaging in early Parkinsonrsquosdiseaserdquo Parkinsonism and Related Disorders vol 10 no 6 pp375ndash379 2004
[3] J L Cummings C Henchcliffe S Schaier T Simuni AWaxman and P Kemp ldquoThe role of dopaminergic imaging inpatients with symptoms of dopaminergic system neurodegen-erationrdquo Brain vol 134 no 11 pp 3146ndash3166 2011
[4] A Siderowf A Newberg K L Chou et al ldquo[ 99mTc]TRODAT-1 SPECT imaging correlates with odor identification in earlyParkinson diseaserdquo Neurology vol 64 no 10 pp 1716ndash17202005
8 Computational and Mathematical Methods in Medicine
[5] Y-H Weng T-C Yen M-C Chen et al ldquoSensitivity andspecificity of 99mTc-TRODAT-1 SPECT imaging in differenti-ating patients with idiopathic Parkinsonrsquos disease from healthysubjectsrdquo Journal of NuclearMedicine vol 45 no 3 pp 393ndash4012004
[6] W-S Huang S-Z Lin J-C Lin S-P Wey G Ting and R-SLiu ldquoEvaluation of early-stage Parkinsonrsquos disease with 99mTc-TRODAT-1 imagingrdquo Journal of NuclearMedicine vol 42 no 9pp 1303ndash1308 2001
[7] P D Acton P T Meyer P D Mozley K Plossl and H F KungldquoSimplified quantification of dopamine transporters in humansusing [ 99mTc]TRODAT-1 and single-photon emission tomog-raphyrdquo European Journal of Nuclear Medicine and MolecularImaging vol 27 no 11 pp 1714ndash1718 2000
[8] P D Acton P D Mozley and H F Kung ldquoLogistic discrimi-nant parametric mapping a novel method for the pixel-baseddifferential diagnosis of Parkinsonrsquos diseaserdquo European Journalof Nuclear Medicine and Molecular Imaging vol 26 no 11 pp1413ndash1423 1999
[9] A J Worth N Makris M R Patti et al ldquoPrecise segmentationof the lateral ventricles and caudate nucleus in mr brain imagesusing anatomically driven histogramsrdquo IEEE Transactions onMedical Imaging vol 17 no 2 pp 303ndash310 1998
[10] V Barra and J-Y Boire ldquoAutomatic segmentation of subcorticalbrain structures in MR images using information fusionrdquo IEEETransactions on Medical Imaging vol 20 no 7 pp 549ndash5582001
[11] Y Xia K Bettinger L Shen and A L Reiss ldquoAutomaticsegmentation of the caudate nucleus from human brain MRimagesrdquo IEEE Transactions on Medical Imaging vol 26 no 4pp 509ndash517 2007
[12] C-F Jiang C-C Chang and S-H Huang ldquoRegions of interestextraction from SPECT images for neural degeneration assess-ment using multimodality image fusionrdquo MultidimensionalSystems and Signal Processing vol 23 pp 437ndash449 2012
[13] A P Dhawan ldquoImage registrationrdquo in Medical Image AnalysisM Akay Ed pp 251ndash276 IEEE Press Piscataway NJ USA2003
[14] A P Dhawan L K Arata A V Levy and J Mantil ldquoIterativeprincipal axes registrationmethod for analysis ofMR-PETbrainimagesrdquo IEEE Transactions on Biomedical Engineering vol 42no 11 pp 1079ndash1087 1995
[15] C-F Jiang C-H Huang and S-T Yang ldquoUsingmaximal cross-section detection for the registration of 3D image data of theheadrdquo Journal of Medical and Biological Engineering vol 31 no3 pp 217ndash226 2011
[16] MKass AWitkin andD Terzopoulos ldquoSnakes active contourmodelsrdquo International Journal of Computer Vision vol 1 no 4pp 321ndash331 1988
[17] L D Cohen and I Cohen ldquoFinite-element methods for activecontour models and balloons for 2-D and 3-D imagesrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol15 no 11 pp 1131ndash1147 1993
[18] B Leroy I Herlin and L Cohen ldquoMulti-resolution algorithmsfor active contour modelsrdquo in Proceedings of the 12th Inter-national Conference on Analysis and Optimization of SystemsImages (ICAOS rsquo96) pp 58ndash65 1996
[19] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998
[20] B Erkol R H Moss R J Stanley W V Stoecker and EHvatum ldquoAutomatic lesion boundary detection in dermoscopy
images using gradient vector flow snakesrdquo Skin Research andTechnology vol 11 no 1 pp 17ndash26 2005
[21] J Tang and S T Acton ldquoVessel boundary tracking for intravitalmicroscopy via multiscale gradient vector flow snakesrdquo IEEETransactions on Biomedical Engineering vol 51 no 2 pp 316ndash324 2004
[22] J Tang S Millington S T Acton J Crandall and S HurwitzldquoSurface extraction and thickness measurement of the articularcartilage fromMR images using directional gradient vector flowsnakesrdquo IEEE Transactions on Biomedical Engineering vol 53no 5 pp 896ndash907 2006
[23] N Tanki K Murase M Kumashiro et al ldquoQuantification ofleft ventricular volumes from cardiac cine MRI using activecontour model combined with gradient vector flowrdquo MagneticResonance in Medical Sciences vol 4 no 4 pp 191ndash196 2005
[24] C Xu and J L Prince ldquoGradient vector flow deformable mod-elsrdquo in Handbook of Medical Imaging pp 159ndash169 AcademicPress 2000
8 Computational and Mathematical Methods in Medicine
[5] Y-H Weng T-C Yen M-C Chen et al ldquoSensitivity andspecificity of 99mTc-TRODAT-1 SPECT imaging in differenti-ating patients with idiopathic Parkinsonrsquos disease from healthysubjectsrdquo Journal of NuclearMedicine vol 45 no 3 pp 393ndash4012004
[6] W-S Huang S-Z Lin J-C Lin S-P Wey G Ting and R-SLiu ldquoEvaluation of early-stage Parkinsonrsquos disease with 99mTc-TRODAT-1 imagingrdquo Journal of NuclearMedicine vol 42 no 9pp 1303ndash1308 2001
[7] P D Acton P T Meyer P D Mozley K Plossl and H F KungldquoSimplified quantification of dopamine transporters in humansusing [ 99mTc]TRODAT-1 and single-photon emission tomog-raphyrdquo European Journal of Nuclear Medicine and MolecularImaging vol 27 no 11 pp 1714ndash1718 2000
[8] P D Acton P D Mozley and H F Kung ldquoLogistic discrimi-nant parametric mapping a novel method for the pixel-baseddifferential diagnosis of Parkinsonrsquos diseaserdquo European Journalof Nuclear Medicine and Molecular Imaging vol 26 no 11 pp1413ndash1423 1999
[9] A J Worth N Makris M R Patti et al ldquoPrecise segmentationof the lateral ventricles and caudate nucleus in mr brain imagesusing anatomically driven histogramsrdquo IEEE Transactions onMedical Imaging vol 17 no 2 pp 303ndash310 1998
[10] V Barra and J-Y Boire ldquoAutomatic segmentation of subcorticalbrain structures in MR images using information fusionrdquo IEEETransactions on Medical Imaging vol 20 no 7 pp 549ndash5582001
[11] Y Xia K Bettinger L Shen and A L Reiss ldquoAutomaticsegmentation of the caudate nucleus from human brain MRimagesrdquo IEEE Transactions on Medical Imaging vol 26 no 4pp 509ndash517 2007
[12] C-F Jiang C-C Chang and S-H Huang ldquoRegions of interestextraction from SPECT images for neural degeneration assess-ment using multimodality image fusionrdquo MultidimensionalSystems and Signal Processing vol 23 pp 437ndash449 2012
[13] A P Dhawan ldquoImage registrationrdquo in Medical Image AnalysisM Akay Ed pp 251ndash276 IEEE Press Piscataway NJ USA2003
[14] A P Dhawan L K Arata A V Levy and J Mantil ldquoIterativeprincipal axes registrationmethod for analysis ofMR-PETbrainimagesrdquo IEEE Transactions on Biomedical Engineering vol 42no 11 pp 1079ndash1087 1995
[15] C-F Jiang C-H Huang and S-T Yang ldquoUsingmaximal cross-section detection for the registration of 3D image data of theheadrdquo Journal of Medical and Biological Engineering vol 31 no3 pp 217ndash226 2011
[16] MKass AWitkin andD Terzopoulos ldquoSnakes active contourmodelsrdquo International Journal of Computer Vision vol 1 no 4pp 321ndash331 1988
[17] L D Cohen and I Cohen ldquoFinite-element methods for activecontour models and balloons for 2-D and 3-D imagesrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol15 no 11 pp 1131ndash1147 1993
[18] B Leroy I Herlin and L Cohen ldquoMulti-resolution algorithmsfor active contour modelsrdquo in Proceedings of the 12th Inter-national Conference on Analysis and Optimization of SystemsImages (ICAOS rsquo96) pp 58ndash65 1996
[19] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998
[20] B Erkol R H Moss R J Stanley W V Stoecker and EHvatum ldquoAutomatic lesion boundary detection in dermoscopy
images using gradient vector flow snakesrdquo Skin Research andTechnology vol 11 no 1 pp 17ndash26 2005
[21] J Tang and S T Acton ldquoVessel boundary tracking for intravitalmicroscopy via multiscale gradient vector flow snakesrdquo IEEETransactions on Biomedical Engineering vol 51 no 2 pp 316ndash324 2004
[22] J Tang S Millington S T Acton J Crandall and S HurwitzldquoSurface extraction and thickness measurement of the articularcartilage fromMR images using directional gradient vector flowsnakesrdquo IEEE Transactions on Biomedical Engineering vol 53no 5 pp 896ndash907 2006
[23] N Tanki K Murase M Kumashiro et al ldquoQuantification ofleft ventricular volumes from cardiac cine MRI using activecontour model combined with gradient vector flowrdquo MagneticResonance in Medical Sciences vol 4 no 4 pp 191ndash196 2005
[24] C Xu and J L Prince ldquoGradient vector flow deformable mod-elsrdquo in Handbook of Medical Imaging pp 159ndash169 AcademicPress 2000
