Artificial intelligence-driven body composition analysis and its association with survival in patients with colorectal liver metastases.

INTRODUCTION: Although computed tomography (CT)-based assessment of body composition has demonstrated that the prognostic value in oncological patients of traditional single-slice manual segmentation is limited by variability and time constraints. This study aims to evaluate the prognostic significance of artificial intelligence (AI)-driven, fully automated volumetric body composition analysis in patients with colorectal liver metastases (CRLM). MATERIAL AND METHODS: Clinical and imaging data were collected from 177 patients with CRLM (105 males and 72 females) with a mean age of 59.72 ±12.19 years. Contrast- enhanced CT scans performed within six weeks of partial hepatectomy were processed using a custom AI-driven segmentation pipeline. Segmentation accuracy was assessed using the Dice coefficient. A Cox proportional- hazards model was employed to analyse the relationship between body composition parameters and overall survival. RESULTS: The segmentation model achieved a median Dice coefficient above 0.99. The final survival model identified nine significant predictors of overall survival, including muscle segment volume percentage, mean Hounsfield units (HU) of the muscle segment, maximum tumour size (cm), sex, presence of multiple metastases (≥ 2), extrahepatic disease, prior chemotherapy before liver resection, non-alcoholic steatohepatitis, and histopathological treatment response > 50%. A higher muscle volume percentage was associated with improved survival (HR = 0.69, p < 0.05), while increased mean HU in muscle segment related to a higher hazard of death (HR = 1.36, p < 0.05). CONCLUSIONS: Automated skeletal muscle measurements revealed significant associations with survival, emphasising their role in outcome prediction and underscoring the need for further validation in larger, multi-institutional cohorts.