MultiprotocolProstateMRIImageAnalysis:ImageSegmentation,Registration,andComputeraidedDiagnosis
Anant Madabhushi1, Jonathan Chappelow1, Satish Viswanath1, Pallavi Tiwari1, Rob Toth1, Mark Rosen2, Michael Feldman2, John E
Tomaszewski2,ElizabethGenega3,NeilRofsky3,RobertLenkinski3,andNicolasBloch3
Laboratory for Computational Imaging and Bioinformatics (LCIB), Department of Biomedical Engineering, Rutgers University, 2
HospitalattheUniversityofPennsylvania,3BethIsraelDeaconessMedicalCenter.
Abstract: At the Laboratory for Computational Imaging and Bioinformatics (LCIB) at Rutgers University and in
collaborationwithclinicalinvestigatorsattheUniversityofPennsylvaniaandBethIsraelDeaconessMedicalCenterwe
have been developing registration, segmentation and computeraided diagnosis (CAD) tools for detection of prostate
cancer (CaP) from multiprotocol MRI including MR Spectroscopy (MRS), T2w, and dynamic contract enhanced (DCE)
MRI.Belowwebrieflydescribe4relatedprojectscurrentlybeingpursuedatLCIB.
A.MultimodalImageRegistrationforDeterminingCaPExtentonMRI:MultimodalregistrationofhistologyandMRIis
critical for CAD evaluation of CaP [1]. Our registration technique termed Combined FeatureEnsemble Mutual
Information(COFEMI) providestheabilitytorobustlyregisterimagesfromverydifferent modalities,suchas MRI and
histology, as well as images that are affected by severe artifacts. COFEMI utilizes highlevel texture feature
representationsoftheimagestoincorporateadditionalinformationintothesimilaritymeasureandachieverobustand
accuratealignment. InFigure1a prostatewholemounthistologyimage(Fig.1(a))isregistered toa correspondingin
vivoMRIimage(Fig.1(b))usingCOFEMIfollowedbythinplatesplinesnonlinearimagewarping.Theresultinghistology
image(Figure1(c))fromwhichthegroundtruthforCaPextentismappeddirectlyontotheMRI(Figure1(d)).
(a)
(b)
(c)
(d)
Figure1:Registrationof(a)wholemounthistologyto(b)exvivoMRIusingCOFEMIandthinplatesplines,resultsin(c)deformed
histologythatisinalignmentwith(b)MRI.Thecancergroundtruthforisthenmappedfrom(c)to(d).
B.ComputeraideddiagnosisofprostatecancerfromhighresolutionT2wandDCEMRI:Ourcurrentworkinprostate
cancerdetectionhasfocusedondevelopingunsupervisedtechniquesforinvivoendorectalT2wandDCEMRIdatavia
application of manifold learning and clustering techniques [2]. Having corrected the MRI data for image intensity
artifacts,theT2wMR(Figure2(a))imageryischaracterizedintermsofalargenumberoftexturefeatures,including2nd
order statistical (Haralick) and gradient (Gabor). The result of these operations is a high dimensional (400+) texture
featurespace.ThisdataisthenanalyzedviamanifoldlearningtechniquessuchasLocallyLinearEmbeddingandGraph
Embedding, which project such high dimensional data into a lower dimensional subspace, while preserving object
adjacencies. The data is then classified in the lower dimensional subspace via unsupervised consensus kmeans
clustering. We have found impressive results upon application of this methodology to T2w and DCEMRI data with
sensitivityandspecificityvaluesofover80%onover301.5Teslaand3TeslainvivoMRIstudies.
(a)
(b)
(c)
(d)
(e)
(f)
(g)
Figure2.(a)OriginalT2wMRimagewithpotentialregionofcanceroverlaidinred.(b)Resultofvisualizinglowerdimensional
subspace onto this image. Regions with similar colors are adjacent in the subspace (c) Final CAD result (via unsupervised
classification)correspondingtoCaPshowninred.Notethecorrespondencetothesurrogategroundtruth(outlinedinwhite).
(d)SampleDCEMRIslicefromtimepoint4(e)Groundtruthestimateofcancermappedontothesliceviaregistration(f)Result
ofvisualizinglowerdimensionalembeddingontotheslice(g)Finalresultofunsupervisedclassificationbestcorrespondingto
cancer shown in red. Note the correspondence to the ground truth in (e).
C.ComputeraideddiagnosisofprostatecancerfromMagneticResonanceSpectroscopy:OurworkinCADforMRSfor
CaPdetectionemploystheuseofnonlineardimensionalityreduction(NLDR)methodstotransformhighdimensional
spectraldatatoanalternatelowerdimensionalspace[3].ClassificationofeachMRspectraisobtainedviaunsupervised
kmeans clustering which clusters low dimensional representations of MR spectra (obtained from NLDR) into distinct
classesbasedonspectraladjacencyinreduceddimensionalspace.Wehaveobtainedhighdetectionaccuracyofcloseto
87%usingournovelintegratedNLDRclusteringschemecomparedtothecurrentstateofartMRSanalysistechniques
like peak detection, zscore and principal component analysis (PCA) (Figure 3). To qualitative validate the clustering
results obtained from our scheme; we perform Independent Component Analysis (ICA) on each cluster where each
independentcomponent(IC)iscomparedtoatypicalCaP,benignsignature(Figure3(e)(g)).Figures3(e)(g)showthat
theICobtainedfromtheclusteridentifiedasCaPviaourscheme(3(f))correspondsresemblestypicalCaPMRspectra
(3(e))comparedtotheICobtainedviaPCA(3(g)).
(a)
(b)(c)(d)
(e)(f)(g)
Figure3:ClusteringresultsusingpopularMRSanalysisschemes(a)peakdetection,(b)zscoreand(c)PCAcomparedto(d)our
NLDR scheme. The white box superposed on the T2 corresponds to the ground truth region. In each image the red cluster
correspondstothelocationsidentifiedascancerbyeachofthemethods(bluevoxelsbeingthoseidentifiedbythemethodas
benign).(g)showsatypicalCancerMRspectra,(h)showstheICobtainedfromclustersidentifiedascancerbyourmethod,(g)
showsthecorrespondingICobtainedviaPCAasDRmethod.Notethehighcorrelationof(e)and(f)comparedto(g).
References
1.Chappelow,J,Madabhushi,A,Rosen,M,Tomasezweski,J,Feldman,M,MultimodalImageRegistrationofexvivo4TeslaProstate
MRIwithWholeMountHistologyforCancerDetection,SPIEMedicalImaging,vol.6512(1),pp.S1S12,2007.
2.Vishwanath,S.,Madabhushi,A.,Rosen,M.,ManifoldLearningandConsensusClusteringforSegmentationof3TeslaProstate
MRI,SPIEMedicalImaging,vol.6915(1),2008.
3.Tiwari,P,Madabhushi,A,Rosen,M,AHierarchicalUnsupervisedSpectralClusteringSchemeforDetectionofProstateCancerfrom
MagneticResonanceSpectroscopy(MRS),MICCAI2007,vol.4792,pp.27886.
