Unraveling Patterns in mTLE: A DWI-Centric and Machine Learning Investigation.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Epilepsy Research

Background:

• Mesial temporal lobe epilepsy (mTLE) is a common drug-resistant epilepsy.
• Diffusion weighted imaging (DWI) offers a radiation-free alternative to techniques like 18F-FDG PET for mTLE lateralization.

Purpose of the Study:

• To develop and evaluate machine learning models utilizing DWI-derived features for classifying left mTLE, right mTLE, and healthy controls.
• To compare the efficacy of different feature selection and classification algorithms for mTLE lateralization.

Main Methods:

• Collected DWI data from 66 subjects (24 left mTLE, 22 right mTLE, 20 controls).
• Extracted features using MRtrix software and compared three feature selection methods (genetic algorithm, PCA, XGBoost).
• Classified subjects using four algorithms (SVM, decision tree, ridge classifier, naive Bayes) with 5-fold cross-validation.

Main Results:

• The genetic algorithm proved superior for feature selection.
• The ridge classifier achieved high accuracies: 0.957 (left vs. normal), 0.957 (right vs. normal), and 0.839 (left vs. right).
• Key discriminative features included local efficiency, modularity, clustering coefficient, betweenness centrality, and PageRank.

Conclusions:

• DWI-based machine learning models show significant potential for automated mTLE lateralization.
• This non-invasive approach can aid clinical decision-making in mTLE diagnosis and surgical planning.
• DWI combined with AI presents an effective neuroimaging strategy for identifying the affected brain hemisphere in mTLE.