Uncertainty-Aware Triage of Microsatellite Instability Status in Colorectal Cancer from H&E-Stained Whole-Slide Images

Microsatellite instability (MSI) is a key biomarker in colorectal cancer (CRC). Accurate distinction between MSI and microsatellite stable (MSS) tumors is critical for diagnosis, prognosis, and therapeutic decision-making. We developed an uncertainty-aware deep learning model for MSI prediction directly from H&E-stained whole-slide images. The approach is based on attention-based multiple instance learning, which can learn from slide-level labels. We extended the approach and included ensembling for more robust prediction. For training, we used 1,492 slides from three public archives and for evaluation 1,094 slides from five independent external cohorts. The approach provides predictions with confidence intervals and rejects uncertain cases. It achieved sensitivity over 97% on 59-78% of the cases, depending on the cohort. These results show that MSS cases can be identified with high confidence while uncertain cases can be rejected to avoid unreliable predictions, enabling clinically reliable MSS triage in colorectal cancer with AI.