Preoperative Prediction of Colorectal Cancer T Stage Using Radiomics and Deep Learning on Non-contrast CT Images: A Dual-Center Study

Objective To establish a combined model integrating deep learning, radiomics features, and relevant clinical characteristics based on Non-contrast CT images. This approach aims to predict colorectal cancer T staging (T1-2 vs. T3-4) through a simple, non-invasive examination method, thereby improving patient prognosis and contributing to the design of personalized therapy plans. Method This retrospective study included 261 patients from two independent centers. Deep learning features were extracted from pre-trained convolutional neural networks (Inception V3, ResNet50, VGG16) using transfer learning.Three machine learning algorithms were utilized to establish three types of models: the traditional radiomic model, the deep learning model, and the radiomics-deep learning combined model (DLR). The optimal model was interpreted using Shapley Additive exPlanations (SHAP). Additionally, univariate and multivariate logistic regression analyses were performed to identify independent clinical risk factors and establish a clinical model. A combined model was developed by integrating these clinical characteristics with the selected deep learning and radiomic features using logistic regression,and a nomogram was created for visual interpretation. Model performance was evaluated by receiver operating characteristic (ROC) analysis, calibration with calibration curves, and clinical utility with decision curve analysis. Results Model performance comparisons revealed that the deep learning radiomics combined model based on the SVM algorithm demonstrated optimal performance, surpassing traditional radiomics models, deep learning models, and clinical models. It achieved an AUC of 0.933 on the training set, 0.865 on the internal validation set, and 0.864 on the external validation set. Incorporating clinical features further improved performance to an AUC of 0.942 on the training set, 0.874 on the internal validation set, and 0.872 on the external validation set. Calibration curves and decision curves demonstrated the combined model's excellent fit. Conclusion The combined model, which combines clinical, radiomics, and deep learning features, exhibits superior predictive performance, serving as a simple, non-invasive, and practical tool for predicting colorectal cancer T staging.