Imaging and Computing Seminar

Bruce Fischl , Harvard Medical School

Computational Modeling of Neuroanatomy using MRI

Computational neuronanatomy can either refer the manner in which the computational architecture of the brain helps it to carry out computations, or the application of computational techniques to build models of neuroanatomical structures. While the former definition is the one that is of ultimate interest, it most likely requires the models indicated in the latter definition. In this talk I will discuss research at MGH with the goal of building such models. The neuronanatomical structures of interest can be broadly subdivided into two categories - cortical and non-cortical.

Cortical structures (particularly the cerbral cortex) are typically highly folded, thin sheets of gray matter. Functionally, the cerebral cortex has been shown to have a "columnar" architecture. For this reason, we construct surface-based models for analysis of cortical properties. The construction of such models is a difficult task due to the high degree of folding of the cortical manifold in conjunstion with the limited (~ 1 mm) resolution of current neuroimaging technologies. Once constructed, the cortical models can be deformed for morphometry, visualization and registration purposes. I will show some results of this type of analysis, including the morphometric changes that the cortex undergoes in disorders such as schizophrenia, Alzheimer's disease, and Huntington's disease, as well as healthy aging.

A different set of techniques have been developed for the construction of models of subcortical structures. Here, we build on work of Kapur, Grimson and Wells, and model the segmentation as an anisotropic nonstationary Markov Random Field. The anisotropy lets us model the local spatial relationships that exist between neuroanatomical structures (e.g. hippocampus is anterior and inferior to amygdala), while the nonstationarity facilitates the encoding of inhomogeneous properties of the tissue within a structure. This approach is based on extracting the relevant model parameters from a manually labeled training set, and has been shown to be comparable in accuracy to the manual labeling.