Beyond BMI: Deep Learning Segmentation-Driven CT Reveals Body Composition Changes after Metabolic and Bariatric Surgery.

Abstract

BMI is the primary metric used to evaluate outcomes of metabolic and bariatric surgery (MBS), but it does not distinguish tissue compartments or quantify visceral adiposity (VAT), a key determinant of cardiometabolic risk. We evaluated the relationship between BMI and VAT and characterized compartment-specific remodeling after MBS using artificial intelligence-enabled CT segmentation. A retrospective analysis of prospectively collected abdominal CT scans was performed at a single tertiary center. Images were processed using Comp2Comp, a validated deep learning pipeline for automated segmentation of visceral adipose tissue, subcutaneous adipose tissue, and skeletal muscle. A population cohort of 435 adults with BMI greater than or equal to 25 kg/m 2 undergoing CT for clinical indications was analyzed to assess baseline BMI-VAT associations. A longitudinal MBS cohort (n = 39 with complete follow-up; 151 CT studies; follow-up to 89 months) was evaluated for temporal changes in BMI, VAT, and muscle. In the population cohort, VAT was moderately correlated with BMI ( r  = 0.36, p < 0.001); however, this association was not significant among patients with BMI greater than or equal to 35 kg/m 2 ( r  = 0.10, p = 0.37). In the MBS cohort, BMI and VAT demonstrated a weak correlation ( R2  = 0.237, p < 0.001). Postoperatively, VAT reduction showed a stronger temporal association ( R2  = 0.661, p < 0.001) than BMI decline ( R2  = 0.571, p < 0.001), with BMI plateauing at longer follow-up. Skeletal muscle demonstrated a distinct recovery trajectory ( R2  = 0.551, p < 0.001). BMI incompletely reflects postoperative tissue remodeling, particularly sustained VAT reduction, after MBS. Artificial intelligence-enabled CT volumetric analysis demonstrates proof of concept for compartment-specific assessment beyond BMI. Prospective validation is required to determine whether VAT-derived metrics more accurately predict cardiometabolic outcomes.