The Taub Faculty of Computer Science Events and Talks

We propose a non-parametric probabilistic model for the automatic segmentation of medical images. The resulting inference algorithms register individual training images to the new image, transfer the segmentation labels and fuse them to obtain the final segmentation of the test subject. Our generative model yields previously proposed label fusion algorithms as special cases, but also leads to a new variant that aggregates evidence locally in determining the segmentation labels. We demonstrate the advantages of our approach in two clinical application: segmentation of neuroanatomical structures and segmentation of the left heart atrium whose shape varies significantly across the population.