BME MS Defense: Sharmistha Chaudhuri

Automatic Segmentation of Ventricles and Outer Brain Contour in Human Brain MR Images using Level Set Model

Supervised by Prof. Prof. Walter O'Dell

Abstract

A random walk model has been developed of tumor cell migration using Magnetic Resonance Diffusion Tensor Imaging (MR-DTI) for predicting the location of secondary tumor/recurrences and based on the hypothesis that the paths of elevated water diffusion along white matter tracts provide a preferred route for migration of tumor cells. The tumor cells are constrained to move only in the brain and cannot enter the ventricles or cross the sulci or gyri, and this is implemented as the boundary condition for the model. Currently, the Brain Extraction Tool in FSL is commonly used for such a purpose, but due to anomalies in patient brains compared to normal brain images, BET often fails on the data sets that we are most interested in studying.

The main aim of this work is to automatically segment the ventricles and the outer brain region of human brain MR images with greater accuracy than conventional methods thereby enabling us to obtain a complete brain mask. A variety of image segmentation techniques exist and have been applied previously to extract features of interest from digital images including thresholding methods, boundary-based methods, region-based methods and hybrid approaches. However, the complexity of the 3D structures in the brain along with variability in anatomy between individuals, disruption of normal anatomy due to presence of large tumors and/or surgical intervention and MRI related noise and inhomogeneity have thwarted the accurate automatic segmentation of anatomic structures using these traditional methods.

In this work, I propose a novel algorithm that performs nearly automatic segmentation of the outer brain and the ventricles in human brain MR images based on the Mumford-Shah level set model and a semi-automated 3D flood-fill operation. The implicit level set function approach to track moving interfaces is especially useful for segmenting structures that have topological changes such as branching or merging of surfaces. To further extract the brain mask alone, we have applied a semi-automated 3D flood-fill operation in the brain mask region. The algorithm was trained on MR image data sets obtained from two glioma patients and tested on four patient data sets. To measure the performance of the image segmentation algorithm, the results were compared to a ground truth database that was obtained unbiased by manual segmentation by an expert individual. The results of the comparison show a dramatically improved match between the ground truth and the image segmentation results, establishing the utility of the algorithm that was developed. The programming languages and software tools used in this work were BET, ImageJ, ITK-SNAP, Java and MATLAB.