Brain MRI Segmentation using a Modified Spectral Clustering Algorithm

Magnetic Resonance Images (MRI) of the brain are invaluable tools to help physicians diagnose and treat various brain diseases including stroke, cancer, and epilepsy. With respect to other biomedical imaging modalities, the MRI technique is far superior at imaging soft tissue. This is because MR image contrast is formed directly from the magnetic properties of water molecules in their local chemical environment and the density of water in soft tissue is high compared to other tissues such as bone. Segmentation of brain MRI’s is an important image processing procedure for both the physician and the brain researcher. For the physician, segmentation may automatically provide an accurate parsing of normal and potentially pathological tissue which is important in clinical diagnosis and treatment monitoring. For the researcher, image segmentation is important for a number of reasons, some of which include visualization of the surface contours of deep brain structures and coregistration of anatomical scans with functional scans. This later problem is often encountered in the functional brain imaging community. Image segmentation is a classic problem in the machine learning and pattern recognition literature. Perhaps the most popular techniques are central grouping techniques such as K-means or Gaussian mixture model fitting using EM [5]. These techniques are computationally efficient but are limited by their Gaussian assumptions about the data. There are many segmentation problems where these assumptions are not valid including MRI, where the data traditionally have complicated, spatially varying statistical properties. Recently, however, a promissing new technique based on pairwise grouping called spectral clustering has emerged in the machine learning community [6, 8, 3, 7]. The central idea of this technique is to find the principle