Development of a whole-slide-level segmentation-based dMMR/pMMR deep learning detector for colorectal cancer

To investigate whole-slide-level prediction in the field of artificial intelligence identification of dMMR/pMMR from hematoxylin and eosin (H&E) in colorectal cancer (CRC), we established a segmentation-based dMMR/pMMR deep learning detector (SPEED). Our model was approximately 1,700 times faster than that of the classification-based model. For the internal validation cohort, our model yielded an overall AUC of 0.989. For the external validation cohort, the model exhibited a high performance, with an AUC of 0.865. The human‒machine strategy further improved the model performance for external validation by an AUC up to 0.988. Our whole-slide-level prediction model provided an approach for dMMR/pMMR detection from H&E whole slide images with excellent predictive performance and less computer processing time in patients with CRC. [Display omitted] •SPEED is a segmentation-based method used to detect dMMR/pMMR from WSIs of CRC•SPEED achieved excellent predictive performance and less computer processing time•A human-machine fusion strategy further improved external validation performance Health sciences; Medicine; Oncology; Health technology