Handcrafted Versus Deep Feature Extraction Methods for MRI-Based Multiple Sclerosis Diagnosis

Background: Despite significant advances in medical image analysis, automated diagnosis of Multiple Sclerosis (MS) from magnetic resonance imaging (MRI) remains challenging due to the complexity of 3D brain data and the variability of lesion appearance. Objective: In this work, we propose an efficient and optimized feature extraction framework for automated MS diagnosis using FLAIR, T1-, and T2-weighted MRI. The approach enhances Decimal Descriptor Patterns (DDP) by integrating local gradient information, producing a 3D texture representation that is more discriminative and expressive. Methods: The study is divided into two main parts: (i) detection of MS, and (ii) assessment of disease progression in affected patients. In each part, features are extracted from the relevant MRI data and classified using multiple classical machine learning classifiers, including Linear Discriminant Analysis (LDA), Support Vector Machines (SVM), k-Nearest Neighbors (KNN), and Logistic Regression. Furthermore, the performance of the proposed handcrafted feature-based approach was compared to features extracted using a deep learning-based model (vision-language model, VLM), specifically CLIP (Contrastive Language-Image Pretraining), enabling a clear comparison of their performance. To assess robustness and generalizability, two complementary validation strategies were adopted: (i) controlled experiments on the BrainWeb dataset under varying T1/T2 contrast conditions, and (ii) validation on a the real-world FLAIR MRI dataset, reflecting clinically relevant lesion visibility. Results: Gradient-DDP features achieve the best overall performance for MS progression, reaching up to 97% accuracy on T2-weighted MRI with SVM, while LDA and Logistic Regression also remain strong with accuracies around 83-96% on T2. For binary MS detection, the proposed method attains near-perfect results, with up to 99% accuracy on FLAIR (SVM/KNN) and 98% on T2-weighted images across SVM, while other classifiers also maintain high performance above 90%. Conclusions: Gradient-DDP provides strong consistency and transparency, offering an interpretable link between texture patterns and diagnostic outcomes. While VLM features perform well when lesion patterns are clearly defined (e.g., in T2), Gradient-DDP demonstrates greater robustness in more challenging modalities such as Flair, where deep representations may be less stable.