CT radiomics plus pathomics nomogram predicts microsatellite instability in colorectal cancer with high accuracy

Multiomics deep learning combining computed tomography (CT) radiomics, pathology features, and clinical variables predicted microsatellite instability in colorectal cancer with AUCs of 0.996, 0.999, and 0.993 in training, internal, and external validation (n=509, 2 centers).

• Retrospective multicenter study of 509 patients with colorectal cancer; center 1 (n=379) split into training/internal validation, and center 2 (n=130) served as external validation.
• Multiphase CT and hematoxylin-eosin pathology features were extracted using a pretrained ResNet-101 network; component models had AUCs from 0.919 to 0.963, while the clinical model alone achieved AUCs of 0.790, 0.761, and 0.756 across the three sets.
• The multiomics nomogram integrated CT radiomics, pathomics, and clinical variables and was externally validated, with interpretation supported by SHAP.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.