基于双编码器与扩散模型的三重对比学习的脑疾病诊断OA

This paper proposed CL-MambaGIN,a triple contrastive learning framework combining dual encoders with a diffu-sion model,to address the high annotation costs in resting-state fMRI analysis and insufficient data augmentation in traditional contrastive learning.The method employed a graph isomorphism network encoder to extract spatial features of brain functional networks,while integrating a Mamba temporal encoder to capture dynamic properties of BOLD signals.The framework intro-duced a diffusion model to generate physiologically realistic augmented samples and implements triple contrastive learning for cross-dimensional feature alignment.Experimental results on autism and depression diagnosis across multiple sites show superior classification performance compared to baseline models.The results demonstrate that the framework effectively improves brain disorder diagnosis through spatiotemporal feature fusion,reduces overfitting in small-sample scenarios,and enhances generalization to unseen data sites.