BrainGraphDiff: A framework for enhanced brain network analysis via adaptive subgraph generation

The diagnosis and identification of mental disorders are widely recognized as a global challenge. The emergence of Graph Neural Networks (GNNs) has significantly advanced brain network analysis. However, the heterogeneity and scarcity of medical imaging data present substantial challenges in terms of generalization and robustness, limiting the reliability and scalability of brain network analysis models in practical clinical applications. One potential solution is to enhance the diversity of training data through generative models, thereby addressing issues such as prediction bias and poor generalization resulting from data scarcity. Building on this, we propose the BrainGraphDiff framework, which incorporates a partial graph generation module to significantly optimize the prediction results of brain network analysis models while ensuring efficiency. However, a key challenge lies in selecting a unified subgraph extraction strategy for all samples, given the heterogeneity of the data. To address this, we introduce the GL-PGIB strategy, which adaptively adjusts the subgraph extraction scope related to labels by using key graph structures as anchors. Our experimental results demonstrate that we can customize subgraph extraction for different samples, and the subsequent generative module effectively balances model efficiency and performance.