CT Radiomics and Ensemble Learning for 5-Year Survival Prediction in Colorectal Liver Metastases

Abstract

Colorectal liver metastases (CRLM) significantly impact patient survival with high recurrence rates. Traditional prognostic models often overlook tumor heterogeneity, leading to suboptimal risk stratification. To address this, radiomics was employed to quantify sub-visual tumor phenotypes, while ensemble learning was selected to robustly handle high-dimensional feature complexity and improve generalization capability. This retrospective study analyzed 145 CRLM patients from The Cancer Imaging Archive, extracting 1130 radiomics features from preoperative CT scans alongside clinical variables. Data were split into training (n=101) and testing (n=44) sets, with feature selection reducing the input to 16 key features. Three ensemble models (XGBoost, LightGBM, Random Forest) were optimized using Optuna, incorporating SMOTE and isotonic calibration. On the test set, XGBoost achieved ROC-AUC 0.918, sensitivity 0.739, and specificity 0.952. LightGBM yielded ROC-AUC 0.916, sensitivity 0.782, and specificity 0.904. Random Forest recorded ROC-AUC 0.888, sensitivity 0.826, and specificity 0.667. Key features included "progression or recurrence" and wavelet-based texture metrics reflecting tumor heterogeneity. These findings demonstrate the effectiveness of combining CT radiomics with gradient boosting models to capture complex prognostic patterns. This integration enhances 5-year survival prediction in CRLM, offering a non-invasive tool for personalized risk stratification and improved clinical decision-making compared to the currently utilized traditional prognostic models.