There are currently no deep learning models applying resting-state functional magnetic resonance imaging (rs-fMRI) data to distinguish patients with Parkinson's disease (PD) and healthy controls (HCs). Moreover, no study has correlated objective gait parameters with brain network alterations in patients with PD. We propose BrainNetCNN + CL, applying a convolutional neural network (CNN) and joint contrastive learning (CL) method to brain network analysis to classify patients with PD and HCs, and compare their performance with classical classification methods. This study aimed to explore more accurate abnormal connecting regions that may serve as potential therapeutic targets, and to explore the correlation between abnormal connecting regions and gait parameters.We enrolled 29 patients with PD and 38 HCs. Rs-fMRI data and high-resolution three-dimensional structural T1-weighted images were acquired for each participant. BrainNetCNN + CL were utilized to classify the PD and HC groups.The top 20 connections with the highest contribution to the classification results obtained using BrainNetCNN + CL included the default mode network (DMN), ventral attention network (VAN), and limbic network (LN). The strength of the functional connectivity (FC) between the right inferior occipital gyrus and left postcentral gyrus in the PD group was negatively correlated with the step length at the self-selected pace (SSP) speed in the "ON" state (P=0.001, r=-0.589). The strength of the FC between the right fusiform gyrus and the right calcarine fissure and surrounding cortex was negatively correlated with the Beck Anxiety Inventory (BAI) score (P=0.032, r=-0.406) and positively correlated with the Berg Balance Scale (BBS) score measured in the "ON" state (P=0.037, r=0.395).BrainNetCNN + CL accurately identified abnormally connected regions associated with gait impairments, which may serve as potential therapeutic targets for PD.2025 AME Publishing Company. All rights reserved.