Global and Regional Deep Learning Models for Multiple Sclerosis Stratification From MRI.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Prognosis

Background:

• Anatomical MRI combined with deep learning (e.g., CNNs) shows promise for predicting multiple sclerosis (MS) prognosis.
• Lack of studies comparing different input strategies for these predictive models.

Purpose of the Study:

• To compare whole-brain versus regional/specific-tissue input strategies for deep learning models.
• To stratify MS patients based on disability level using MRI data.

Main Methods:

• Retrospective study with 319 MS patients (in-house) and 440 MS patients (external validation).
• 3D-CNN trained to classify patients by Expanded Disability Status Scale (EDSS) score (≥3.0 vs. <3.0).
• Compared global (whole brain) and regional (white matter, gray matter, etc.) input strategies.

Main Results:

• In-house: Gray matter regional model achieved 81% accuracy; global approach achieved 79%.
• External validation: White matter model achieved 72% accuracy; global approach achieved 71%.
• Global approach showed the best performance-to-generalization trade-off.

Conclusions:

• The global approach provides a robust method for stratifying MS patients by disability using anatomical MRI and deep learning.
• Input strategy choice impacts model performance in predicting MS prognosis.