Automated Body Composition from Computed Tomography Scans Improves Survival Prediction in Colorectal Cancer Patients

Colorectal cancer is one of the most common cancers, but the current staging system is limited by the variability and paradoxical survival outcomes. Body composition is an accurate predictor of survival and can be extracted from routine computed tomography (CT) scans used in cancer diagnosis. However, despite its potential, the adoption of body composition analysis has been limited due to the challenges in generating the data. In this study, we propose a deep learning-based model that combines clinical and body composition biomarkers to predict the survival of colorectal cancer patients. Our best model, which integrates both clinical and body composition features, achieved a time-dependent concordance-index score of 0.7298 ( $p<0.001$ ), demonstrating a significant improvement over models based solely on clinical or body composition biomarkers, indicating that models combining body composition and clinical markers could improve survival prediction. Additionally, we observed that increased skeletal muscle tissue area and radiodensity were associated with reduced mortality risk, while higher radiodensities of visceral and subcutaneous adipose tissues were associated with increased risk.