A Unified Graph Domain Adaptation Framework for Cross-Site Brain Network Analysis

Compared to single-site brain network analysis, multi-site methods leverage larger sample sizes to achieve more robust prediction and generalizability, which has garnered increasing attention in brain disease diagnosis. However, these approaches also face heightened heterogeneity across sites (e.g., scanner protocols) and subjects (e.g., clinical variability), along with potential outliers with distributional shifts in connectivity patterns that may misguide model training. To address these challenges, we propose a unified graph domain adaptation framework for cross-site brain network analysis, which mitigates the heterogeneity at the sample and site levels while preserving their topological structure. Specifically, we first develop graph convolution to extract the topological representations of brain networks in the source and target sites. Next, adversarial learning is employed to learn domain invariant representations, while domain shift across sites is mitigated. Additionally, we further align intra-class similarity and inter-class dissimilarity across sites by contrastive learning, enhancing feature discriminability and alleviating sample heterogeneity. Finally, we incorporate a self-paced learning strategy that gradually assesses the difficulty and uncertainty of samples, which guides the model to prioritize simpler samples during the early training, thereby mitigating the influence of difficult samples and further reducing sample heterogeneity. Extensive experiments on multi-site brain disorder datasets demonstrate that our proposed method outperforms state-of-the-art methods in cross-site brain disease analysis tasks.