Domain adversarial transfer graph deep learning: A cross-site MDD fMRI data analysis framework

The pathological mechanisms of depression are not yet fully clear, and the causative factors remain somewhat ambiguous. Clinical diagnosis of depression is often challenging and complex, frequently leading to misdiagnosis and missed diagnoses. Combining deep learning with resting-state fMRI can quantify the degree of abnormal brain function caused by depression and automatically screen for discriminative features that aid in the classification and identification of depression, which may serve as hypothesis-generating discriminative features within the current dataset, providing candidate neuroimaging signatures that warrant further investigation. This paper proposes a cross-site fMRI data analysis framwork for depression. First, it contains a graph deep learning-based auxiliary diagnostic method, which fully leverages the topological structure of brain networks to achieve higher classification accuracy compared to existing models, with interpretable results. Building on this network, the framework also contains a domain-adversarial-based cross-site semi-supervised transfer method is proposed, making full use of multi-site data to analyze depression-related brain networks and ROIs. Finally, based on cross site data, the distribution of brain networks and brain regions was discussed. The research findings are consistent with existing studies, confirming the reliability of this method. Furthermore, we validated the cross-dataset generalizability of our framework on an independent OpenNeuro dataset, where adversarial transfer consistently outperformed direct transfer, demonstrating the potential of our approach to generalize beyond the original consortium.