Asian Research Thesis Index

Abstract

This thesis aims to understand low level segmentation and quantification aspects of brain MR slices purely from imaging perspective. In this context we propose two fully automatic and novel preprocessing techniques for improvement in low level segmentation. At first place we propose a novel method for cerebrum localization which reduces extraneous information from brain MR slices significantly. Up to our best knowledge no one used this cerebrum extraction approach before. Secondly a novel polygonal seed selection procedure is suggested and preferred over histogram peaks method in the absence of prior to improve initial conditions for region growing. In addition to this we present a generalized conceptual framework for region growing segmentation designed through exhaustive region growing literature review. Extensive qualitative results have been shown over full brain MR volume for segmentation. The quantification aspect is also attempted to compare segmentation results with available ground truth. Ten T1-Weighted (voxel size: 1 mm3; dimensions: 181 x 217 x 181) Normal brain phantom datasets with varying noise and inhomogeneity along-with true anatomical model have been downloaded from McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University and are subject to low level segmentation using Otsu’s method, Seeded Region Growing Method (SRG), Watershed Transformation and K-Means Clustering. The preprocessing strategies with respect to brain MR images include intensity normalization, local histogram smoothing, non-local mean (NLM) filtering and cerebrum localization. To demonstrate cerebrum extraction over full brain MR volume the Otsu’s method along with hill down optimization is used which neatly separates the four tissue classes in brain MR slices, namely Background (BG), Cerebrospinal fluid (CSF), Gray Matter (GM) and White matter (WM). This initial Otsu’s segmentation along with novel polygonal seed selection scheme puts global information into service to improve initial conditions for seeded region growing (SRG). The polygonal seeds idea is also experimented with watershed segmentation and K-Means clustering of brain MR images showing clear improvement in initial conditions but more work needs to be done with geometrical foreground markers and seeds. Confusion matrix analysis shows that Otsu’s algorithm fails to segment CSF pixels especially when noise level is increased. In comparison to Otsu’s segmentation region growing always identify CSF pixels with an accuracy of around 30% to 50% over entire brain MR volume. Otsu’s method identifies GM and WM pixels most of the time with an accuracy of 88% to 99% percent while region growing capacity to classify these pixels ranges from 60% and 90% respectively. Normalized Root Mean Square Error (NRMSE) for CSF, GM and WM volume densities of ten subjects comes out 1.1%, 3.3% and 3.3% respectively from Otsu’s Segmentation while it is 3.8%, 6.2% and 5.1 % from region growing results for single subject. In terms of volume density these normal brain phantom data sets contain highest GM density, then WM density and least volume is occupied by CSF voxels.

Add/Update Thesis