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

Area of Science:

• Medical Image Analysis
• Neurology
• Machine Learning

Background:

• Automated Multiple Sclerosis (MS) diagnosis from MRI is challenging due to complex 3D brain data and variable lesion appearance.
• Existing methods struggle with the nuances of MS lesion visualization across different MRI sequences.

Purpose of the Study:

• To develop an efficient feature extraction framework for automated MS diagnosis and progression assessment using FLAIR, T1, and T2-weighted MRI.
• To enhance the Decimal Descriptor Patterns (DDP) by integrating local gradient information for improved 3D texture representation.

Main Methods:

• Proposed a Gradient-DDP feature extraction method, integrating local gradient information into DDP for enhanced texture analysis.
• Classified features using Linear Discriminant Analysis (LDA), Support Vector Machines (SVM), k-Nearest Neighbors (KNN), and Logistic Regression.
• Compared Gradient-DDP performance against a deep learning vision-language model (VLM), specifically CLIP, on controlled (BrainWeb) and real-world (FLAIR) datasets.

Main Results:

• Gradient-DDP achieved up to 97% accuracy for MS progression assessment on T2-weighted MRI using SVM.
• For binary MS detection, the method reached near-perfect accuracy, up to 99% on FLAIR (SVM/KNN) and 98% on T2-weighted images.
• Gradient-DDP demonstrated superior robustness on challenging FLAIR images compared to VLM features.

Conclusions:

• Gradient-DDP offers a transparent and interpretable approach for MS diagnosis and progression monitoring via MRI texture analysis.
• The proposed handcrafted feature method shows high accuracy and robustness, particularly in challenging imaging conditions where deep learning models may falter.