Prototypical Representation Learning for Multi-Site Domain Generalization in Schizophrenia Diagnosis

Brain functional networks (BFNs) derived from multi-site fMRI data have been widely used in the diagnosis of psychiatric disorders such as schizophrenia (SZ). However, site-induced distribution shifts pose a major challenge to the generalization of classification models based on BFNs. While domain generalization (DG) seeks to address this challenge without access to target domain data, most methods make the idealized assumption that each class exhibits consistent structures across sites. The inherent diversity within each class often violates this assumption, resulting in poor generalization to unseen domains. To address this issue, we proposed a DG framework based on a transformer encoder combined with prototype learning. We first employed a transformer encoder to capture global dependencies among brain regions and extract discriminative subject-level representations from BFNs. A site-independence module with Hilbert-Schmidt Independence Criterion (HSIC) regularization was then applied to enforce site-invariant feature learning. Projected features were softly assigned to multiple prototypes via Sinkhorn matching, with prototypes updated via exponential moving average (EMA), and the maximum likelihood estimation (MLE) loss further refining feature-to-prototype matching probabilities. Contrastive and alignment losses were applied to promote inter-class separability and intra-class consistency, respectively. Experimental results showed that our method outperformed the other 7 DG, 6 domain adaptation (DA), 6 multi-site and 6 state-of-the-art methods across two independent SZ datasets (88.89%±2.22% and 86.05%±1.64%). Ablation results highlighted the contributions of the MLE, contrastive, and alignment constraints to the performance improvement. Furthermore, the identified discriminative temporal regions provided insights into the dysfunctional neural-mechanisms in SZ.