Deep learning-based mismatch repair prediction using colorectal cancer macroscopic images: a diagnostic study

Mismatch repair (MMR) testing is recommended for all colorectal cancer (CRC) patients, but this assay necessitates the involvement of specialized institutions and is time-consuming. This study aims to develop a deep learning model for MMR prediction using macroscopic images to provide a rapid and cost-free screening tool.

This diagnostic study enrolled 809 CRC patients who underwent surgical resection without neoadjuvant therapy at Peking University Third Hospital (from January 2020 to July 2025). Macroscopic images of surgical specimens were captured immediately after resection. MMR status was confirmed by postoperative immunohistochemical staining for MMR proteins (MLH1, MSH2, MSH6, and PMS2). Deep learning models were developed by a two-step approach: automated lesion segmentation using DeepLabV3 + , followed by MMR classification using vision transformer (ViT). MMR prediction performance was mainly evaluated utilizing area under the curve (AUC). Gradient-weighted Class Activation Mapping (Grad-CAM) appraisal and principal component analysis (PCA) were performed to assess the explainability of the model.

The proposed model achieved an average AUC of 0.896 (95% CI, 0.763-0.959) on internal test and 0.860 (95% CI, 0.644-0.921) on independent test for MMR prediction. High NPVs of 0.987 (95% CI, 0.928-0.999) and 0.978 (95% CI, 0.925-0.994) were observed in internal and independent testing, respectively, using a threshold of 0.323. Grad-CAM analysis and PCA demonstrated that the deep-learning model was of explainability.

The new deep-learning model accurately identified MMR status using macroscopic specimen images and showed potential for MMR screening among CRC patients, particularly in a rapid-response scenario.