A Modified Expectation-Maximization Approach for HMRF-Based Brain MRI Classification

Accurate segmentation of brain tissues in magnetic resonance imaging (MRI) is essential for clinical diagnosis, pathological assessment, prognosis evaluation, and brain development studies. However, tissue heterogeneity resulting from bias field distortion, partial volume effects, noise, and magnetic field inhomogeneities poses significant challenges. In this study, we propose a Hidden Markov Random Field model combined with a Modified Expectation-Maximization algorithm (HMRF-EM) to improve segmentation accuracy by accounting for neighborhood correlation and signal intensity non- uniformity. The algorithm was implemented in R and evaluated on T1-weighted simulated Brain Web data. The model effectively segmented cerebrospinal fluid (CSF), gray matter (GM), and white matter (WM) with tissue proportions of 35%, 47%, and 18%, respectively. Validation results demonstrated a mean square error of 0.0290, misclassification rate of 0.0870, and tissue volume errors of 0.0578 (CSF), 0.0246 (GM), and 0.0063 (WM). Dice similarity coefficients were 0.9244, 0.9086, and 0.9134 for CSF, GM, and WM, respectively. These findings indicate that the proposed HMRF-EM approach yields reliable and accurate brain tissue classification, making it suitable for clinical and research applications.