Explainable 3D CNNs link regional and network level disruption in early Parkinson's MRIs to symptom progression

Parkinson's Disease (PD) is a progressive neurodegenerative disorder affecting approximately 1% of the population over 65. Clinical diagnosis typically depends on tracking gradually developing motor symptoms as the disease progresses, underscoring the need for early detection methods to aid intervention while symptoms are still minor. Inexpensive and widely available imaging modalities such as T1-weighted MRI (T1w MRI) have potential for early PD diagnosis but lack established systematic biomarkers of PD pathology. In this study, a 3D convolutional neural network (3D CNN) was trained on 100 predominately early-state PD and 100 control T1w MRIs from Parkinson's Progression Markers Initiative (PPMI), achieving a classification accuracy of 84.5%. Misclassified subjects were majority unmedicated and particularly early PD (< 3 years since first symptoms). To interrogate the biological basis behind the model's decisions, novel explainability methods were applied to generate regional saliency maps from both PD and control classifications. Regional saliency across subjects correlated best with cognitive and motor scores in nigrostriatal and other subcortical regions, as well as in temporal and insular cortices, indicating changes in these areas were best connected with symptom progression. The model was also sensitive to changes in the left frontal cortex across many subjects, which exhibited the greatest raw saliency magnitude. Pairwise saliency correlation was most pronounced between areas within the same functional network, suggesting the CNN was sensitive to network level changes in structural MRI. These findings demonstrate the potential of explainable 3D CNNs to identify network and regional biomarkers of early PD from T1w MRI.