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

Parkinson's Disease (PD) affects 1-2% of adults over 65 and is diagnosed by tracking motor symptoms, underscoring the need for earlier detection. T1-weighted MRI (T1w MRI) is inexpensive and widely available but can lack the resolution to detect established markers of PD pathology, such as substantia nigra degeneration. We trained 3D convolutional neural networks (CNNs) on five site-stratified folds of 249 T1w MRIs from 141 early PD subjects (<10 years since first symptoms) and 185 from 126 controls. Because every fold was evaluated only on sites absent from its training data, the ensemble accuracy of 67.8% (AUC 0.670) reflects generalization to unseen sites. To interrogate these decisions, we parcellated saliency from a cohort-contrastive GradCAM, which baselines PD-class saliency against controls, and from Integrated Gradients, using the Julich Brain atlas. Both showed saliency in neocortical areas matching later Braak stages (5-6), when pathology reaches the neocortex, and lower subcortical saliency, suggesting limited sensitivity to earlier stages. Saliency was more correlated within functional networks than among random region groupings, and pooled saliency tracked motor symptoms and medication rather than cognitive scores or duration. These findings show explainable 3D CNNs can identify meaningful network and regional signatures of early PD from T1w MRI.