Domain Adaptation Method Based on Multi-Scale Brain Network Feature Fusion for Brain Diseases Diagnosis

Brain networks are essential for human cognition and crucial in diagnosing brain diseases. Traditional single-site methods, limited by small and homogeneous samples, struggle to generalize. In contrast, multi-site methods enhance diagnostic accuracy by integrating diverse data but face domain shifts due to inter-center data differences. To mitigate this issue, domain adaptation approaches strive to diminish the distributional discrepancies between source and target domains. However, they often overlook the complex topological structures in brain networks. This paper proposes a domain adaptation method based on multiscale brain network feature fusion for multi-site brain disease diagnosis. It utilizes graph convolutional networks to extract brain network features at different scales, thereby fully capturing the complex topological structure information. Moreover, it employs adversarial domain adaptation and prototype learning to mitigate domain and semantic shifts, thereby enhancing model generalization. Our approach significantly outperforms conventional domain adaptation methods in experiments conducted on multi-site Attention Deficit Hyperactivity Disorder (ADHD) datasets.