Retrospective study develops a deep learning model to accurately assess muscle volume in individuals, indicating improved efficiency in body composition analysis.
Key Points
The aim is to create and evaluate a deep learning model for automated measurement of trunk muscle volume using PET/CT scans.
Developed an nnU-Net-based segmentation model using data from 20 manually annotated datasets.
Segmented trunk muscles in a sample of 209 individuals undergoing PET/CT and bioelectrical impedance analysis.
Validated model performance using Dice similarity coefficients and compared muscle mass assessments with bioelectrical impedance analysis.
The model achieved a Dice similarity coefficient of 0.991.
Automated 3D muscle volume showed a strong correlation with BIA-derived mass (r = 0.961; 95% CI: 0.948, 0.970) compared to 2D metrics (r = 0.912; 95% CI: 0.886, 0.933; P < 0.001).
2D L3 cross-sectional area demonstrated reduced performance in men (r = 0.757; P < 0.001 vs. 3D volumetry).