Gaussian mixture model - Expectation maximization algorithm for brain images

Abstract

Segmentation of human brain can be performed with the aid of mathematical algorithm as well as computer-based system to assist radiologists and medical related profession to monitor the condition of one's brain comprehensively. Due to the complex structure of the human brain, one cannot simply analyze them just by looking at the MRI images. This research examines the brain segmentation and the validation of the segmentation using ground truth data for seven subjects. The segmentation of brain regions such as white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF) can be accomplished by using Gaussian Mixture Model (GMM) and Expectation-Maximization (EM) Algorithm. The results of segmentation are shown by the Gaussian distribution graph that indicates the volume of brain regions. The segmentation results are validated by the value of Dice index, Jaccard index, and positive predictive value (PPV). It is found that all seven subjects have high value for every index as the values ranging from more than 0.6 to almost approaching 1. For all subjects, the lowest percentage for Dice is 77.82% while the highest is 84.28%, the lowest percentage for Jaccard is 63.70% while the highest is 72.84%, and the lowest percentage for PPV is 94.44% while the highest is 98.75%. In conclusion, the index values for all subjects are acceptable and this means the segmentation by using GMM and EM Algorithm is accurate after going through the process of validation of segmentation.