Can Multiomics Modeling Enable Accurate Prediction of Microsatellite Instability in Colorectal Cancer?
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiology, Tongde Hospital of Zhejiang Province, Hangzhou, Zhejiang, China (W.A., G.M., S.D.); Zhejiang Academy of Traditional Chinese Medicine, Hangzhou, Zhejiang, China (W.A.).
- Zhejiang Chinese Medical University, Hangzhou, Zhejiang, China (Y.H.).
- Department of Radiology, Tongde Hospital of Zhejiang Province, Hangzhou, Zhejiang, China (W.A., G.M., S.D.).
- Department of Pathology, Tongde Hospital of Zhejiang Province, Hangzhou, Zhejiang, China (W.W.).
- Department of Radiology, Taizhou First People's Hospital, School of Medicine, Taizhou University, Taizhou, Zhejiang, China (D.H., Y.Z.).
- Department of Radiology, Tongde Hospital of Zhejiang Province, Hangzhou, Zhejiang, China (W.A., G.M., S.D.). Electronic address: [email protected].
Abstract
To investigate a multimodal deep learning model integrating computed tomography (CT) radiomics and histopathologic pathomics for predicting microsatellite instability (MSI) in colorectal cancer (CRC). A total of 509 patients with pathologically confirmed CRC were retrospectively enrolled from two medical centers. Patients were divided into an MSI group and a microsatellite stable group. Patients from center 1 (n = 379) were randomly divided into training (n = 261) and internal validation (in-vad) sets (n = 118); center 2 (n = 130) served as the external validation (ex-vad) set. Deep learning features were extracted from multiphase CT and hematoxylin-eosin-stained images using a pretrained ResNet-101 network. After feature selection optimized by Monte Carlo simulation, a CT deep learning radiomics score (DLRS) and a pathomics score (DLPS) were developed. In addition, a preoperative prediction model integrating clinical variables and the DLRS was constructed. These were combined with significant clinical variables into a multiomics nomogram. Model performance was evaluated using receiver operating characteristic (ROC) analysis, calibration curves, and decision curve analysis, with interpretation via SHapley Additive exPlanations. Among the developed models, the multiomics nomogram demonstrated the best performance, with areas under the ROC curve (AUCs) of 0.996, 0.999, and 0.993 in the training, in-vad, and ex-vad sets, respectively. The AUCs for the DLRS, DLPS, and preoperative models ranged from 0.919 to 0.963, while the clinical model yielded AUCs of 0.790, 0.761, and 0.756 across the three sets. The integrated deep learning nomogram incorporating CT radiomics, pathomics, and clinical variables enables accurate prediction of MSI status in CRC, showing potential as a noninvasive tool for molecular stratification. A multiomics deep learning framework integrating multiparametric CT radiomics, histopathologic pathomics, and clinical variables enables highly accurate prediction of microsatellite instability in colorectal cancer. Developed and validated using datasets from two medical centers, this strategy offers a promising imaging-derived biomarker for molecular stratification and personalized treatment decision-making in patients with colorectal cancer.