Deepath-MSI: a clinic-ready deep learning model for microsatellite instability detection in colorectal cancer using whole-slide imaging
Feng Xu, Wenjuan Yin, Qing Ye, Yayun Chi, Huer Wen, Yifeng Sun, Jin Zheng, Qifeng Wang, Qian Wang, Ming Zhao, Yuan Lin, Qinghua Xu, Dan Su, Xiaoyan Zhou
Abstract
Microsatellite instability (MSI) is crucial for immunotherapy selection and Lynch syndrome diagnosis in colorectal cancer. Despite recent advances in deep learning algorithms using whole-slide images, achieving clinically acceptable specificity remains challenging. In this large-scale multicenter study, we developed Deepath-MSI, a feature-based multiple instances learning model specifically designed for sensitive and specific MSI prediction, using 5070 whole-slide images from seven diverse cohorts. Deepath-MSI achieved an AUROC of 0.98 in the test set. At a predetermined sensitivity threshold of 95%, the model demonstrated 92% specificity and 92% overall accuracy. In a real-world validation cohort, performance remained consistent with 95% sensitivity and 91% specificity. Deepath-MSI could transform clinical practice by serving as an effective pre-screening tool, substantially reducing the need for costly and labor-intensive molecular testing while maintaining high sensitivity for detecting MSI-positive cases. Implementation could streamline diagnostic workflows, reduce healthcare costs, and improve treatment decision timelines.