ProTA: Combining Prototypical Transformer and Prototype-Based Aggregation for Microsatellite Instability Classification on Whole-Slide Images

Whole-slide image (WSI) classification primarily relies on multiple instance learning (MIL). However, due to the heterogeneity of tumor tissue and the severe imbalance between positive and negative instances within bags, existing models often suffer from limited representational capacity in generating slidelevel embeddings. This limitation hampers classification accuracy and restricts model generalization. To address these issues, we propose ProTA, a framework that combines a Prototypical Transformer and a Prototype-Based Aggregation mechanism for the classification of Microsatellite Instability (MSI) in WSIs. Specifically, the Prototypical Transformer dynamically calibrates initial prototypes via a self-attention mechanism, capturing the semantic relationships between instances and prototypes. Then, leveraging the optimized prototypes, a Gaussian Mixture Model (GMM) is employed to accurately model the distribution of patch features and estimate mixture parameters, thereby yielding a compact and structured WSI-level embedding. Finally, a structured multi-layer perceptron (MLP) performs slide-level MSI classification. Extensive cross-dataset experiments conducted on TCGA-COAD, PAIP2020, and HCH demonstrate that, compared to mainstream MIL methods (ABMIL, CLAM, TransMIL) and prototype-based models (PAMIL, PANTHER), the proposed approach achieves the best ACC and AUC across all three datasets, validating its effectiveness and superiority in the MSI classification task. The source code is available at https://github.com/Sitara3337/ProTA.