A Mixture-of-Experts Decision Support System for Digital Pathology

Whole slide image (WSI) classification is a core task in digital pathology that can assist decision-making procedures for pathologists. Several models, mainly built based upon multiple-instance learning, have shown to be effective in processing and analyzing WSIs. However, these are designed, trained, and evaluated on a single classification task, and thus the models are limited to a specific task and cannot utilize the data and knowledge from other tasks. This substantially limits the ability and expandability of the model to support clinical decision-making. In this study, we present a mixture-of-experts decision support system for digital pathology. The proposed decision support system merges multiple individual models, of which each is tailored to a specific task, and forms a unified model equipped with group intelligence that can handle multiple classification tasks. The proposed system utilizes Transformer architecture to process WSIs and a language decoder to enable flexible classification across multiple tasks. The experiments were conducted on five datasets: CAMELYON16, TCGA-BRCA, TCGA-NSCLC, TCGA-RCC, and TCGA-ESCA, achieving accuracies of 96.124%, 95.062%, 90.805%, 95.402%, and 91.071%, and F1 of 0.965, 0.971, 0.908, 0.711, and 0.918, respectively. These results demonstrate the effectiveness of the proposed approach in supporting clinical decision-making.