BioCompNet: A Deep Learning Workflow Enabling Automated Body Composition Analysis toward Precision Management of Cardiometabolic Disorders.

Abstract

Growing evidence highlights the importance of body composition (BC), including bone, muscle, and adipose tissue (AT), as a critical biomarker for cardiometabolic risk stratification. However, conventional methods for quantifying BC components using medical images are hindered by labor-intensive workflows and limited anatomical coverage. This study developed BioCompNet-an end-to-end deep learning workflow that integrates dual-parametric magnetic resonance imaging (MRI) sequences (water/fat) with a hierarchical U-Net architecture to enable fully automated quantification of 15 biomechanically critical BC components. BioCompNet targets 10 abdominal compartments (vertebral bone, psoas muscles, core muscles, subcutaneous AT [SAT], superficial SAT, deep SAT, intraperitoneal AT, retroperitoneal AT, visceral AT, and intermuscular AT [IMAT]) and 5 thigh compartments (femur, muscle, SAT, IMAT, and vessels). The workflow was developed on 8,048 MRI slices from a community-based cohort (n = 503) and independently validated on 240 MRI slices from a tertiary hospital (n = 30). The model's performance was benchmarked against expert annotations. On internal and external validation datasets, BioCompNet achieved average Dice similarity coefficients of 0.944 and 0.938 for abdominal compartments and 0.961 and 0.936 for thigh compartments, respectively. Excellent interreader reliability was observed (intraclass correlation coefficient ≥ 0.881) across all quantified features, and IMAT quantification showed a strong linear trend (P _trend < 0.001) compared to physician-rated assessments. The workflow substantially reduced processing time from 128.8 ± 5.6 to 0.12 ± 0.001 min per case. By enabling rapid, accurate, and comprehensive volumetric analysis of BC components, BioCompNet establishes a scalable framework for precision cardiometabolic risk assessment and clinical decision support.