FocusedON-BC: A Robust Deep Learning Framework for Automated Body Composition Assessment

Abstract

Background: Computed tomography-based body composition assessment enables the quantification of clinically relevant prognostic conditions such as sarcopenia, myosteatosis, and visceral adiposity; the manual segmentation process limits its routine implementation in clinical practice. We developed FocusedON-BC, an automated deep learning tool for opportunistic screening of skeletal muscle (SM), visceral adipose tissue (VAT), and subcutaneous adipose tissue (SAT) across the T12–L5 range; Methods: Validated on a multicenter cohort of 518 patients (3280 slices) with diverse body mass index (12.7–47.7 kg/m2) from different computed tomography manufacturers. Performance was benchmarked against expert segmentation using the Dice coefficient score (DSC) and the mean absolute error (MAE); Results: FocusedON-BC achieved expert-level accuracy: mean DSC was (Formula presented.) (SM), (Formula presented.) (VAT), and (Formula presented.) (SAT). Clinical MAE remained <5% for all compartments. Performance was robust, independent of body mass index and computed tomography scanner model. Qualitative assessment confirmed the tool’s capability to isolate intermuscular adipose tissue for radiodensity analysis; Conclusions: FocusedON-BC provides accurate, vendor-agnostic body composition and muscle quality analysis. Its reliability across diverse phenotypes supports implementation for routine nutritional screening.