Explainable 3D CNNs Yield Progression-Linked Profiles of Saliency in Early Parkinson's Disease Using T1-Weighted MRI

Explainable 3D CNNs trained on site-stratified T1-weighted MRI distinguish early Parkinson’s disease from controls with 67.8% ensemble accuracy (AUC 0.670) across unseen sites.

• Why it matters: Earlier detection is needed because PD diagnosis relies on motor symptom tracking, while T1w MRI may not resolve established pathology markers like substantia nigra degeneration.
• What they did: The model used five site-stratified folds of 249 early PD scans (<10 years since first symptoms) and 185 control scans, with GradCAM and Integrated Gradients saliency mapped to the Julich Brain atlas.
• The result: Saliency patterns aligned with later Braak stages (5–6) in neocortical areas and showed lower subcortical saliency, with saliency linked to motor symptoms and medication rather than cognitive scores or duration.

AI summary of the abstract · may contain errors · not medical advice · read the paper.