A novel approach to MRI Brain Tumor delineation with Independent Components & Finite Generalized Gaussian Mixture Models

Automated segmentation of tumors from a multispectral data set like that of the Magnetic Resonance Images (MRI) is challenging. Independent Component Analysis (ICA) and its variations for Blind Source Separation (BSS) have been employed in previous studies but have met with cumbersome obstacles due to its inherent limitations. Here we have approached the multispectral data set initially with feature extraction followed by a kernel shape based unsupervised classification method, Finite Generalized Gaussian Mixture Model (FGGM) ICA-FGGM model, for an improved classification of brain tissues in MRI. First, ICA is applied to MRI brain data from 3 source image sets T1, T2 and PD/ FLAIR images to get optimally feature extracted three independent components. FGGM model can then incorporate various distributions from peaked ones to flat ones; thereby overriding the disadvantages of conventional approaches trying to represent data using a single probability density function. ExpectationMaximization algorithm is used to estimate the model parameters. Experiments were carried out initially on synthetic image sets to validate the algorithm and then on normal and abnormal clinical multispectral MRI brain images. Comparative studies using quantitative and qualitative analysis against conventional approaches confirm the effectiveness and superiority of the proposed method.