Brain Tissue Classification of Magnetic Resonance Images Using Conditional Random Fields

In this project, I propose the application of a discriminative framework for segmentation of a T1weighted magnetic resonance image(MRI). The use of Gaussian mixture models (GMM) is fairly ubiquitous in processing brain images for statistical analysis. This generative framework makes several assumptions that restricts its success and application. GMM assumes there is no spatial correlation when classifying tissue type, and also assumes each class of tissues is described by one Gaussian distribution. While the model is simple and quick, both assumptions are often violated leading to poor performance. Conditional Random Fields (CRF) avoid these assumptions by learning to discriminate between different tissue classes without assumptions on the class conditional density distribution. Additionally, CRFs allow arbitrary features of the data to be instantiated, allowing spatial correlations to be added in addition to other rich features other than voxel intensity. Results show that future work on the model is needed to enhance the model and improve the quality of training data