NA-MIC National Alliance for Medical Image Computing Statistical Models of Anatomy and Pathology Polina Golland
National Alliance for Medical Image Computing Statistical Models of Anatomy Applications –Spatial priors for segmentation –Population studies Traditional approach –Align images to a common template –Compute mean and co-variation Challenges –Spatial variability in the structure of interest –Loss of detail –Heterogeneous populations
National Alliance for Medical Image Computing Our Solutions Use training data in novel ways –handle spatial variability TBI, tumors –avoid the loss of detail Atrial Fibrillation, Huntington’s, Alzheimer’s Model heterogeneous populations – capture broader variability Atrial fibrillation, radiation therapy, Alzheimer’s
National Alliance for Medical Image Computing Spatial Priors and Pathology Augmented generative model –Atlas: spatial prior for healthy tissues –Estimate: spatial prior for tumor Output –Common healthy tissue segmentation –Modality-specific tumor segmentation Menze, MICCAI 2010
National Alliance for Medical Image Computing Spatial Priors and Pathology (cont’d) More accurate than EM-segmentation with outlier detection Comparable to within-rater variability Going forward: TBI Menze, MICCAI 2010
National Alliance for Medical Image Computing Label Fusion Segmentation Test Image Subject Specific Label Prior New Segmentation Pairwise Registration Training Data
National Alliance for Medical Image Computing Generative Model for Label Fusion {Ln}{Ln} {In}{In} L(x)I(x) M Test image Training images … … ? Sabuncu, TMI 2010
National Alliance for Medical Image Computing Left Atrium Segmentation More accurate than baseline methods Correctly identified all veins Local prior for scar location Weighted fusion Majority Manual Parametric Mdepa, MICCAI Workshop 2010
National Alliance for Medical Image Computing Modeling Heterogeneous Populations Manifold of anatomical images –Spectral embedding –Statistical model in new space –Gerber, MedIA 2010 Collection of sub-populations –Mixture model –Templates represent population –Sabuncu TMI 2009 noise
National Alliance for Medical Image Computing Applications for Spatial Priors Identify relevant “neighborhood” for the new image –A (small) set of training examples –A (local) atlas template Construct patient-specific spatial prior –Average or use label fusion Challenges: –Reduce the number of pairwise registration steps –Model influence of selected neighborhood on new image
National Alliance for Medical Image Computing Conclusions Clear need for new methods –Handle spatial variability of pathology –Handle anatomical variability in a population Preliminary results: local models –In the image coordinates –In the space of images Going forward –Development in the context of the DBPs