Deep learning and body composition model for predicting postoperative complications in colorectal cancer

Colorectal cancer (CRC) is the third most commonly diagnosed malignancy worldwide and remains a leading cause of cancer-related mortality. Surgical resection remains the cornerstone of curative treatment for CRC; however, postoperative complications, including anastomotic leakage, infections, and thromboembolic events, continue to substantially affect patient prognosis. These complications are associated with prolonged hospitalization, increased healthcare expenditures, delayed postoperative recovery, and elevated mortality risk. Therefore, the accurate identification of patients at high risk for postoperative complications is of considerable clinical importance for optimizing perioperative management and improving surgical outcomes.

A total of 154 patients were retrospectively enrolled, including 99 patients in the training cohort and 55 patients in the external validation cohort. Abdominal CT images at the L1-L5 levels were automatically segmented using a pretrained DeepLabv3-ResNet101 model implemented on the Onekey platform, followed by manual correction to ensure segmentation accuracy. Visceral adipose tissue, subcutaneous adipose tissue, skeletal muscle, and intramuscular adipose tissue were segmented, and body composition indices (VAT_h, SAT_h, SMA_h, and IMAT_h) were subsequently calculated after normalization by patient height. Deep learning features were extracted from CT images using a pretrained 3D ResNet-18 model. Feature selection and model development were conducted independently. Random forest, logistic regression, and extremely randomized trees combined with recursive feature elimination were applied for feature selection. Subsequently, logistic regression, random forest, support vector machine (SVM), k-nearest neighbors (KNN), XGBoost, and LightGBM algorithms were constructed and compared for predictive performance.

Univariate and multivariate analyses identified BMI, SMA_h, SAT_h, and VAT_h as independent risk factors for postoperative complications. Among all predictive models, the combined model incorporating the deep learning score (DL-score) achieved the best performance, with AUC values of 0.944 and 0.855 in the training and validation cohorts, respectively. DeLong test results further demonstrated that both the deep learning model and the combined model significantly outperformed the clinical and body composition models, whereas the combined model provided superior discrimination in identifying high-risk patients.