This repository contains the implementation of an integrated framework for Multiple Sclerosis (MS) lesion segmentation and classification of abnormalities in the Sensory, Motor, and Visual systems. The framework fuses MRI-derived texture features with clinical metadata to enhance the precision of MS lesion detection and classification.
U-Net Attention SqueezeNet (UASNet): A specialized U-Net model incorporating Attention and Squeeze-and-Excitation (SE) blocks for high-precision segmentation of MS lesions on T2-weighted MRI images, achieving 98.82% segmentation accuracy. Feature Extraction: Extraction of 21 texture features from segmented MRI images combined with clinical data for comprehensive analysis. Integration of Imaging and Clinical Data: This approach aims to facilitate more accurate and personalized diagnosis and treatment strategies for MS. Feature Correlation: Pearson, Spearman's Rank, and Kendall's Tau correlation methods were used to analyze relationships between features. Feature Selection: Application of Chi-Square feature selection to improve classification performance. Machine Learning Classifiers: Traditional classifiers like Decision Trees, Random Forests, Naive Bayes, and Logistic Regression were employed for evaluation. Graph-Based Learning: GCN, GAT, GraphSAGE, VGAE, DGI & GIN were tested classifying motor, visual & sensory system abnormalities.