Predicting survival among colorectal cancer patients using automated body composition

Abstract

Colorectal cancer is among the most prevalent cancers worldwide, yet its current staging system is limited by variability and paradoxical survival outcomes. Body composition analysis from routine computed tomography (CT) scans offers predictive markers of patient survival, such as sarcopenia and adiposity. However, manual segmentation of CT images is labor-intensive and prone to inter-observer variability. In this study, we validated the accuracy and reliability of automated CT-based segmentation using the Data Analysis Facilitation Suite (DAFS) Express, achieving high concordance with manual segmentation, with mean DICE scores above 96% for all tissue areas and only 0.10% poor cases across a cohort of 5,973 cancer patients. The DAFS Express method also showed strong concordance with manual analysis in mortality association, as indicated by similar hazard ratios, confidence intervals, and p-values. Furthermore, we developed a deep learning model that combines clinical and body composition biomarkers to predict survival outcomes in colorectal cancer patients. Our integrated model achieved a time-dependent concordance index (Concordancetd- index) score of 0.7298 (p < 0:001), significantly outperforming models based solely on clinical biomarkers. Models relying exclusively on body composition features had the lowest Concordancetd-index scores, suggesting that body composition alone may not be a reliable predictor of survival.