MTSK-DA: Interpretable Multicenter Transfer via Discriminative Alignment and Attention-Guided Graph Regularization

Multicenter learning confronts three intertwined challenges - distributional heterogeneity, class ambiguity, and the pursuit of interpretability. We introduce MTSK-DA: Interpretable Multicenter Transfer via Discriminative Alignment and Attention-Guided Graph Regularization, a Takagi-Sugeno-Kang fuzzy framework that cascades a base-auxiliary-target hierarchy to disseminate knowledge while respecting site-specific sample nuances. At each auxiliary center, Discriminative Statistic Alignment minimizes marginal and class-conditional discrepancies. Simultaneously, it repels interclass outliers to create sharper decision boundaries and a calibrated, transferable rule base. The target model is further refined by a Multicenter Attention-Regularized Laplacian Graph that overlays intraclass attraction and interclass repulsion attention signals onto a Laplacian affinity graph, preserving manifold geometry while amplifying semantic margins. Embedding these statistically grounded, attention-aware modules within an interpretable fuzzy backbone unifies clarity with performance. Extensive cross-center evaluations consistently demonstrate that MTSK-DA significantly enhances class separability and accuracy, decisively surpassing state-of-the-art domain-alignment and fuzzy baselines under distributional drift. By fusing discriminative alignment with attention-guided graph regularization inside an interpretable framework, MTSK-DA charts a compelling route toward trustworthy and transferable recognition across heterogeneous centers.