Abstract
To investigate the feasibility of a radiomics model for the detection of bladder invasion (BI) by colorectal cancer (CRC) on CT images. Ninety-six patients with CRC and a suspicion of BI who underwent tumor resection with partial or total cystectomy were reviewed. The 96 patients were randomly assigned to the training dataset (n = 68) or test dataset (n = 28) at a ratio of 7:3. The CT images were reviewed by two experienced radiologists, who provided a CT impression of the invasion of the bladder by CRC. A region of interest (ROI) on the CT images for each case was manually labeled by two radiologists. A radiomics model was constructed using a Categorical Boosting (CatBoost) classifier. The predicted probability by CatBoost was used to evaluate the efficacy of the radiomics model. The areas under the curve (AUCs) of the receiver operating characteristic were compared between the radiomics model and the CT impression. In the training dataset, the AUC of the radiomic model [0.864 (95% CI: 0.778, 0.951)] was significantly greater than that of CT impression [0.678 (95% CI: 0.569. 0.786), P = 0.007]. In the test dataset, the AUC of the radiomic model [0.883 (95% CI: 0.699, 1.000)] was also significantly greater than that of CT impression [0.570 (95% CI: 0.370, 0.770), P = 0.040]. It is feasible to use radiomics models for the prediction of BI by CRC, which might perform better than human radiologists.