Interpretable neural network based on T1-weighted MRI for detecting Parkinson's disease

Motivation: Applying radiological techniques for early diagnosis of Parkinson's disease (PD) is crucial for prognosis of patients. A simple, accurate technique with high patient compliance is still lacking. Goal(s): To explore the feasibility of using visual neural networks based on T1-weighted images for detecting PD. Approach: Patients with PD and healthy controls (HCs) underwent T1-weighted MRI scans. A convolutional neural network (CNN) architecture was developed. Interpretable maps were also drawn to investigate the efficiency of computerized neural networks. Results: The CNN model achieved achieved good classification performance.Interpretable maps highlighted the critical regions for detecting PD. Impact: CNN based on T1-weighted imaging is a reliable and accurate diagnostic tool for detecting PD. Interpretable maps improve the explainability of classification results in clinical applications.