Juan Eugenio Iglesias, PhD Research Staff, Martinos Center for Biomedic al Imaging (MGH), Lecturer in Radiology, Harvard Medical School Research Affiliate, CSAIL, MIT

Abstract

In this talk, I will present our work on Bayesian segmentation methods for brain MRI analysis. Compared with modern deep learning approaches, these methods have two advantages: They enable us to use ex vivo imaging (e.g., ex vivo MRI, histology), of superior resolution and contrast, to build very detailed atlases; and they are agnostic to the MRI contrast of the input scan to analyze, even if multimodal. First, I will introduce a generative model for brain anatomy, and how we can use Bayesian inference and model selection to build computational atlases within the model given a set of manual segmentations, while accounting for the consistency of brain structures across subjects and the number of training examples. Second, I will explain how the probabilistic model can be extended to generate image intensities, and how this new model can also be “inverted” with Bayesian inference to produce automated segmentations. Finally, I will present some results on segmentation of subregions of the hippocampus, amygdala, and thalamus.