Validation of body composition parameters extracted via deep learning-based segmentation from routine computed tomographies

Sarcopenia and body composition metrics are strongly associated with patient outcomes. In this study, we developed and validated a flexible, open-access pipeline integrating available deep learning-based segmentation models with pre- and postprocessing steps to extract body composition measures from routine computed tomography (CT) scans. In 337 surgical oncology patients, total skeletal muscle tissue (SM_total), psoas muscle tissue (SM_psoas), visceral adipose tissue (VAT), and subcutaneous adipose tissue (SAT) were quantified both manually and using the pipeline. Automated and manual measurements showed strong correlations (SM_psoas: r = 0.776, VAT: r = 0.993, SAT: r = 0.984; all P < 0.001). Measurement discrepancies primarily resulted from segmentation errors, anatomical anomalies or image irregularities. SM_psoas measurements showed substantial variability depending on slice selection, whereas SM_total, averaged across all L3 levels, provided greater measurement stability. Overall, SM_total performed comparably to SM_psoas in predicting overall survival (OS). In summary, body composition measures derived from the pipeline strongly correlated with manual measurements and were prognostic for OS. The increased stability of SM_total across vertebral levels suggests it may serve as a more reliable alternative to psoas-based assessments. Future studies should address the identified areas of improvement to enhance the accuracy of automated segmentation models.