Deciphering multiple sclerosis disability with deep learning attention maps on clinical MRI.

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Convolutional Neural Networks (CNNs) offer advanced MRI analysis for neurological conditions.
• CNNs extract subtle image features, surpassing conventional methods in predicting disease course.
• Attention maps from CNNs can illuminate disease mechanisms linked to disability.

Purpose of the Study:

• To evaluate a 3D-CNN model for predicting neurological condition outcomes using brain MRI.
• To compare the CNN model's performance against a logistic regression (LR) model.
• To investigate CNN-derived attention maps for insights into disability accumulation mechanisms.

Main Methods:

• A 3D-CNN model was trained on T1-weighted and T2-FLAIR MRI scans from 319 patients.
• Patient groups were defined by Expanded Disability Status Scale (EDSS) scores (≥3.0 vs. <3.0).
• Layer-wise relevance propagation generated attention maps; model performance was validated on an independent cohort (N=440).

Main Results:

• The CNN model achieved 79% accuracy, outperforming the LR model (77%).
• External validation without retraining yielded 71% accuracy.
• Attention maps highlighted frontotemporal cortex and cerebellum, suggesting distributed damage contributes to disability.

Conclusions:

• CNNs provide a powerful tool for predicting neurological condition progression from MRI data.
• Attention map analysis offers novel insights into the neuroanatomical basis of disability.
• Disease progression and disability accumulation involve complex patterns of central nervous system damage.