NA-MIC National Alliance for Medical Image Computing UNC Shape Analysis Martin Styner, Ipek Oguz Department of CS UNC Chapel Hill Max Jacob Styner
National Alliance for Medical Image Computing Slide 2 UNC Shape Analysis UNC Shape Analysis Toolbox –SPHARM-PDM, Hotelling, permutation, FDR –Local shape analysis via MANCOVA –Shape analysis via discrimination (with MIT) –Collaborations (Utah, GT) –Over 100 downloads of shape tool distribution Enhanced correspondence: Curvature MDL Ongoing Slicer 3 integration Binary Segmentation Volumetric analysis: Size, Growth Shape RepresentationStatistical analysis Local processes
National Alliance for Medical Image Computing Slide 3 Segmentation Spherical Parameterization SPHARM-PDM Hotelling T 2 metric Surface Distance Hypothesis Testing Permutations, FDR, GLM+MANCOVA Representation Preprocessing - Correspondence - Alignment - Scaling Analysis UNC Shape Analysis Toolbox
National Alliance for Medical Image Computing Slide 4 UNC Shape Analysis Toolbox Publications: MICCAI 06 (2x), ISBI 06, SPIE 07, sub. ISBI 07 Comprehensive including visualization Spherical harmonics + PDM Complex shape: Striatum, from subdivision to local shape analysis NAMIC core interaction –1: Parts of analysis with GT, Utah –3: Harvard PNL Caudate studies Paper in preparation
National Alliance for Medical Image Computing Slide 5 Local Shape with Mancova Current analysis only allows direct group comparisons No corrections for age, gender, weight etc No correlation with variables, such as IQ, clinical scores, age, duration of illness etc Work with D Pantazis, USC Test locally and permutation tests for correction 1.General Linear Model fitting (for each x,y,z) 2.MANCOVA model, Wilks’s & Roy’s Lambda 3.Permutation tests over Test statistics Matlab implementation at USC Application to UNC DBP Autism data drives research (correction for gender, age, IQ)
National Alliance for Medical Image Computing Slide 6 Shape Discrimination Shape analysis via discrimination –How to best discriminate 2 groups –Discrimination direction (DD), linear or radial basis function Application –Distance maps: Golland, MedIA 05 –SPHARM-PDM surfaces –Good agreement hypo test and DD magnitude MIT, Kitware MIT, Kitware, UNC Rbf DD (solid) SPHARM Hypothesis
National Alliance for Medical Image Computing Slide 7 MDL Correspondence with Local Features Ipek Oguz, Martin Styner, Tobias Heimann, Guido Gerig Traditional MDL uses position to establish correspondence Not satisfactory for objects with complicated geometry We incorporate local features (e.g. curvature) to improve correspondence Striatum (caudate + nucleus accumbens + putamen ), coloring is spherical parametrization
National Alliance for Medical Image Computing Slide 8 Criteria for Model Validation Compactness –Ability to use a minimal set of parameters Generalization –Ability to describe instances outside of training set: leave one out Specificity –Ability to represent only valid instances of the objects: Distance to closest sample
National Alliance for Medical Image Computing Slide 9 Compactness Cumulative sum of eigenvalues Normalized to [0,1]Not Normalized Global and local alignment removes variability → more compact model.
National Alliance for Medical Image Computing Slide 10 Generalization Calculate leave-one-out PGA Project left out sample into PGA space Compute distance between original sample and PGA space projection Average over all permutations of leave-one-out
National Alliance for Medical Image Computing Slide 11 Generalization Global and local alignment model has smaller distances. Model seems to flatten out after 5 modes (?)
National Alliance for Medical Image Computing Slide 12 Specificity Compute normally distributed random sample in PGA space with eigenvalues as variance Construct m-rep from random weight vector Compute distance to closest training sample
National Alliance for Medical Image Computing Slide 13 Specificity Multi-object model does not flatten out, more modes might be necessary (with more samples)
National Alliance for Medical Image Computing Slide 14 Group Discrimination HDLSS Problem: High-dimensional feature space Low sample size Overfitting Solutions: Use only first few PGA parameters -> hope that these subspace serves well for discrimination Use robust technique for HDLSS group classification
National Alliance for Medical Image Computing Slide 15 Results - I Simple object geometry SPHARM and MDL on pure curvature (CS) perform poorly MDL over Curvature + position (XYZCS) gives results similar to position (XYZ) only
National Alliance for Medical Image Computing Slide 16 Results - II Complex object geometry SPHARM and pure curvature (CS) performs poorly Curvature + position (XYZCS) gives better results than position only (XYZ)
National Alliance for Medical Image Computing Slide 17 Discussion Methodology With compex object geometry –local curvature improves correspondence Choice of particular curvature metric does not have significant effect –Principal curvatures, Gaussian curvature, mean curvature, curvedness, shape index Our framework can be used for any combination of local features: local curvature, cortical thickness, fMRI, DTI, MRA, etc. MICCAI 2007 submission
National Alliance for Medical Image Computing Slide 18 Slicer 3 Integration External modules for all shape analysis tools in UNC pipeline –Individual modules –Visualization tool –No module for MDL Processing possible –Very tedious –Case by case, step by step…
National Alliance for Medical Image Computing Slide 19 Slicer 3 Modules
National Alliance for Medical Image Computing Slide 20 Next: All-In-One tool Batch processing is necessary for shape analysis from a practical viewpoint Top-level tool for whole shape analysis pipeline –GUI: intuitive, end-user in mind, Slicer 3 external module –Specification of input segmentations –Full shape pipeline computation Use of BatchMake for computing Distributed computing with Condor (BatchMake) –Advanced parameters for experts
NA-MIC National Alliance for Medical Image Computing Future Development: Cortical Correspondence Ipek Oguz, Martin Styner – UNC Josh Cates, Tom Fletcher, Ross Whitaker – Utah
National Alliance for Medical Image Computing Slide 22 Main Idea - Cort Corresp Use entropy-based particle system (Cates) for cortical correspondence –Highly convoluted surface Integrate sMRI, DTI, MRA, fMRI –How to combine these data Single, flexible framework for the cortical surface, subcortical structures and cerebellum
National Alliance for Medical Image Computing Slide 23 Finding Correspondence In order to apply the particle method to the cortex, we need to first ‘inflate’ the surface Possible methods: –FreeSurfer –Area preserving surface evolution (Tannenbaum ?, Faugeras ?,..)
National Alliance for Medical Image Computing Slide 24 Integrating Data Structural –Position, curvature, depth to inflated surface Vascular –Distance to closest vessel(s) of certain size –Distance to labeled vessel(s) DTI –Probabilistic connectivity –To given region(s), intra & inter hemispheric –Locally reduced using priors/thresholds Local vascular & connectivity patterns
National Alliance for Medical Image Computing Slide 25 Example 1 Targeting fMRI –better functional correspondence (better sensitivity) in an amygdala-curcuit related task MRA data: distance to closest arterial vessel of minimal size (2mm) DTI data: connectivity to amygdala
National Alliance for Medical Image Computing Slide 26 Example 2 Cortical thickness comparison with better “anatomic” correspondence MRA: distance to major vessels (arterial & venal) DTI: probabilistic connectivity to all major subcortical structures –Connectivity vector –Possibly train & threshold