A Pathology-Interpretable Deep Learning Model for Predicting Microsatellite Instability State in Colorectal Cancer: Validation across Diverse Platforms and Asian Cohorts

Abstract

Background The determination of microsatellite (MS) state plays a vital role in precise diagnosis and treatment of colorectal cancer (CRC). However, the limited availability of medical resources and challenging economic circumstances render MS state testing unattainable for a significant proportion of CRC patients. We propose a novel pathology-interpretable deep learning model to predict the MS state of CRC, with an inclination to validate in the Asian population across multiple cohorts and sequencing platforms. Methods Pathological images, documented MS state and clinical characteristics of 360 CRC patients from the cancer genome atlas together with 782 cases from Chinese hospital were included. Results The model demonstrated notable generalization ability, achieving an AUC of 0.92 in the independent verification cohort and an AUC of 0.93 in the multicenter cohort. We achieved cell nucleus segmentation and image-based cell type identification using the hover-net model to provide the pathology interpretation of the model, unveiling significant predictors of microsatellite instability. Notably, patients classified as microsatellite instability (MSI) by the model exhibited higher progression-free survival rates, supported by follow-up data. Conclusions The model shows great potential for clinical usage, particularly in the Asian population, demonstrating high accuracy across multiple cohorts and MSI detection platforms.