Blind source separation with mixture models - A hybrid approach to MR brain classification

Abstract

The development of automated segmentation approaches, which do not suffer from excessive computational burden and intra- and inter-observer variability, is the holy grail of multispectral MR image classification. A new segmentation approach to the data set of MR brain images using a combination of Independent Component Analysis (ICA) with a generalized version of the popular Gaussian Mixture Model (GMM) for unsupervised classification is proposed to be superior to conventional methods in this paper. We propose to optimize the parameters of the mixture model using a meta-heuristic approach like the Particle Swarm Optimization (PSO) to escape the problem of local traps (maxima or minima). Experiments were carried out initially on a synthetic MR Brainweb image set as proof of concept and subsequently on 152 sets of clinical MR images with T1w, T2w and FLAIR sequences. The major advantage of the proposed algorithm is the increased accuracy of lesion classification - average of 94.79% (±1.7) against 85.85% (±3.1) without ICA. As a result of the incorporation of ICA, the inherent computational overhead was also lowered as evidenced by faster convergence. Comparative studies using quantitative and qualitative analysis against conventional algorithms establish the superiority of the proposed approach.

Keywords: Finite generalized Gaussian mixtures; Image classification; Independent Component Analysis; MRI; Multispectral; Particle Swarm Optimization.