PD-ARnet: a deep learning approach for Parkinson's disease diagnosis from resting-state fMRI.

Abstract

The clinical diagnosis of Parkinson's disease (PD) relying on medical history, clinical symptoms, and signs is subjective and lacks sensitivity. Resting-state fMRI (rs-fMRI) has been demonstrated to be an effective biomarker for diagnosing Parkinson's disease. Approach: This study proposes a deep learning approach for the automatic diagnosis of PD using rs-fMRI, named PD-ARnet. Specifically, PD-ARnet utilizes Amplitude of Low Frequency Fluctuations (ALFF) and Regional Homogeneity (ReHo) extracted from rs-fMRI as inputs. The inputs are then processed through a developed dual-branch 3D feature extractor to perform advanced feature extraction. During this process, a Correlation-Driven weighting module is applied to capture complementary information from both features. Subsequently, the Attention-Enhanced fusion module is developed to effectively merge two types of features, and the fused features are input into a fully connected layer for automatic diagnosis classification. Main results: Using 145 samples from the PPMI dataset to evaluate the detection performance of PD-ARnet, the results indicated an average classification accuracy of 91.6% (95% confidence interval [CI]: 90.9%, 92.4%), precision of 94.7% (95% CI: 94.2%, 95.1%), recall of 86.2% (95% CI: 84.9%, 87.4%), F1 score of 90.2% (95% CI: 89.3%, 91.1%), and AUC of 92.8% (95% CI: 91.1%, 95.0%). Significance: The proposed method has the potential to become a clinical auxiliary diagnostic tool for Parkinson's disease, reducing subjectivity in the diagnostic process, and enhancing diagnostic efficiency and consistency. .