Purpose:
Rapid decision-making is essential in precision medicine for initiating molecular targeted therapy for patients with cancer. This study aimed to extract pathomorphologic features that enable the accurate prediction of genetic abnormalities in cancer from hematoxylin and eosin images using deep learning (DL).

Experimental Design:
A total of 1,657 images (one representative image per patient) of thin formalin-fixed, paraffin-embedded tissue sections from either primary or metastatic tumors with next-generation sequencing–confirmed genetic abnormalities—including BRAFV600E and KRAS mutations, and microsatellite instability high (MSI-H)—that are directly relevant to therapeutic strategies for advanced colorectal cancer were obtained from the nationwide SCRUM-Japan GI-SCREEN project. The images were divided into three groups of 986, 248, and 423 images to create one training and two validation cohorts, respectively. Pathomorphologic feature-prediction DL models were first developed on the basis of pathomorphologic features. Subsequently, gene-prediction DL models were constructed for all possible combinations of pathomorphologic features that enabled the prediction of gene abnormalities based on images filtered by the combination of pathomorphologic feature-prediction models.

Results:
High accuracies were achieved, with AUCs > 0.90 and 0.80 for 12 and 27, respectively, of 33 analyzed pathomorphologic features, with high AUCs being yielded for both BRAFV600E (0.851 and 0.859) and MSI-H (0.923 and 0.862).

Conclusions:
These findings show that novel next-generation pathology methods can predict genetic abnormalities without the need for standard-of-care gene tests, and this novel next-generation pathology method can be applied for colorectal cancer treatment planning in the near future.