Development of a Deep Learning Model for Automated Measurement of Skeletal Muscle Volume in <sup>18</sup>F-FDG PET/CT.
Authors
Affiliations (5)
Affiliations (5)
- Preemptive Medicine and Lifestyle Related Disease Research Center, Kyoto University Hospital, Kyoto, Kyoto, Japan. [email protected].
- Department of Advanced Healthcare Informatics, Hamamatsu University School of Medicine, Hamamatsu, Shizuoka, Japan.
- Department of Real World Data R&D, Graduate School of Medicine, Kyoto University, Kyoto, Kyoto, Japan.
- Preemptive Medicine and Lifestyle Related Disease Research Center, Kyoto University Hospital, Kyoto, Kyoto, Japan.
- Diagnostic Imaging and Nuclear Medicine, Graduate School of Medicine, Kyoto University, Kyoto, Kyoto, Japan.
Abstract
To develop a deep learning-based automated trunk muscle volumetry method using whole-body CT from PET/CT and evaluate its performance against bioelectrical impedance analysis (BIA). In this retrospective study, an nnU-Net-based segmentation model was developed using a three-step iterative refinement strategy with 20 manually annotated datasets. Trunk muscles were segmented in 209 individuals (median age 52 years, IQR 47-60.3; 148 men) undergoing PET/CT and BIA. Model performance was validated using Dice similarity coefficients (DSCs). Pearson's correlation coefficients (r) compared BIA-derived trunk muscle mass with that estimated by automated 3D volume or conventional 2D L3 cross-sectional area (CSA). Williams' t test compared dependent correlations. The segmentation model achieved a DSC of 0.991. Automated 3D muscle volume demonstrated a significantly stronger correlation with BIA-derived mass (r = 0.961; 95% CI: 0.948, 0.970) compared to 2D L3-CSA (r = 0.912; 95% CI: 0.886, 0.933; P < 0.001). While 3D volumetry maintained high correlations in both sexes (r = 0.898 for both), 2D L3-CSA showed significantly reduced performance in men (r = 0.757; P < 0.001 vs. 3D). Automated 3D volumetry provides a significantly more robust assessment of muscle mass than conventional 2D metrics by capturing whole-trunk anatomical variations. This framework enables accurate, scalable body composition analysis directly from routine PET/CT workflows without additional radiation.