Unsupervised Classification for Volume-Based Magnetic Resonance Brain Images

Abstract

Magnetic resonance (MR) image classification generally performs slice by slice in which case training samples are slice-dependent. Each slice requires its own specific training samples and training samples obtained from one slice are not necessarily applicable to another slice. This paper develops a new approach to unsupervised classification for magnetic resonance images which consists of two stage processes. The first stage develops an unsupervised training sample generation process, called unsupervised volume sphering analysis (UVSA). It comprises of three processes, (1) volume-based data sphering, (2) fuzzy cmeans and (3) an iterative Fisher's linear discriminant analysis (IFLDA). The second stage uses the training samples found in the 1st stage to further perform supervised classification on the entire image data set using the training samples generated by UVSA in the 1st stage. This can be accomplished by implementing the IFLDA once again to produce final classification results. Experimental results demonstrate that the UVSA using one set of training samples not only performs as well as those using training samples specifically selected for individual image slices, but also saves significant amounts of radiologists' efforts in selecting training samples and data processing time.