Interpretable deep learning-based hippocampal sclerosis classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Hippocampal sclerosis (HS) is a common cause of epilepsy.
• Accurate diagnosis of HS is crucial for effective treatment.
• Current diagnostic methods can be subjective and time-consuming.

Purpose of the Study:

• To evaluate a deep learning model for classifying hippocampal sclerosis using clinical MRI data.
• To provide visual interpretations for the model's predictions.
• To assess the model's potential for clinical application.

Main Methods:

• Utilized T2-weighted oblique coronal brain MRI images from patients undergoing epilepsy protocols.
• Trained a deep learning model using a dataset of 320 participants (160 no HS, 100 left HS, 60 right HS).
• Employed layer-wise relevance propagation for model interpretability and visualized prediction contributions.

Main Results:

• Achieved 87.5% multiclass and 88.8% binary classification accuracy on the training set validation.
• Attained 91.5% multiclass and 89.76% binary classification accuracy on the independent test set.
• Generated sparse visual interpretations highlighting key MRI voxels contributing to HS classification.

Conclusions:

• The interpretable deep learning model shows promise for clinical adaptation in diagnosing hippocampal sclerosis.
• The model effectively utilizes common MRI data, even with realistic abnormalities.
• Plausible visual interpretations support neurologists in diagnosing HS.