Juan Eugenio Iglesias Gonzalez

Title - Building a high-resolution computational atlas of the whole human brain with histology, deep learning, and Bayesian modeling: open-source implementation and application to population studies.

Abstract:

Widespread neuroimaging packages like FreeSurfer, FSL or SPM enable morphometric, functional, and connectivity studies of the human brain in vivo using MRI. While these packages are frequently updated to stay near the state of the art in terms of methodology, they still rely on computational atlases which are over a decade old using in vivo MRI scans, and which fail to describe the human brain beyond the whole structure level. In this talk, I will present ongoing work to build a computational atlas of the whole human brain at the substructure level, which we intend to integrate with FreeSurfer. First, I will present our initial work on an atlas of the hippocampal subfields using /ex vivo/ MRI, which represented the first step towards building such an atlas of the whole brain. Both the construction of the atlas and its application to automated image segmentation rely on Bayesian inference techniques within a generative framework of imaging data. Next, I will explain the limitations of /ex vivo/ MRI in atlas building, and how we are overcoming them with histology. Unfortunately, histological analysis introduces a new set of problems associated with the geometric distortion caused by sectioning and staining; I will present techniques based on Bayesian modeling and weakly supervised deep learning that we have developed to correct for this distortion. Finally, I will present preliminary results on population studies of the hippocampus, amygdala and thalamus, as well on the  3D reconstruction of the publicly available atlas of the Allen Institute.