Accurate assessment of body composition is essential for understanding human physiology and health risks and developing personalized medical strategies. Traditional approaches, such as single-slice segmentation at the L3 vertebra, often fail to capture the complexities of whole-body tissue distribution, which is influenced by genetics, metabolism, environment, and physiology. To address these limitations, we present a novel methodology to estimate whole-body composition using sub-body anatomical region computed tomography (CT) scans. For this study, a cohort of 101 subjects with plasmacytoma cancer who underwent whole-body (head-to-toe) CT scans was utilized. Using in-house segmentation software, skeletal muscle (SKM), subcutaneous adipose tissue (SAT), visceral adipose tissue (VAT), and intermuscular adipose tissue (IMAT) were quantified. Generally, sub-bodies were either individual regions, such as the chest (CHE), abdomen (ABD), and pelvis (PLV), or combined regions, including chest and abdomen (CHA) and abdomen and pelvis (ABP), as defined according to standard vertebral landmarks. A multivariate linear regression model incorporating demographic and CT-derived features is developed to predict whole-body tissue volumes. Tenfold cross-validation was used to validate the performance of the prediction models, and evaluation metrics included the mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), 95% confidence intervals across folds, and the coefficient of determination (R2). For individual regions, SKM demonstrates strong performance with R2 values from 0.83 to 0.91 and low MAE (e.g., chest 1.14 ± 0.83 L, R2 = 0.913). SAT performs best in the PLV region (R2 = 0.918, MAE = 1.58 ± 1.28 L), while VAT achieves its highest single-region performance in the ABD region (R2 = 0.950, MAE = 0.37 ± 0.35 L). IMAT shows small MAE values (e.g., chest 0.26 ± 0.21 L) but relatively low R2, with the highest predictability obtained in the CHE region (R2 = 0.822). For combined regions, SKM performs best in the CHA region (R2 = 0.924, MAE = 1.01 ± 0.84 L), while SAT improves in the ABP region (R2 = 0.923, MAE = 1.45 ± 1.35 L). VAT shows consistently excellent performance across both combined regions, with similar R2 values (CHA 0.977, ABP 0.981) and low MAE. Broader anatomical coverage improved predictions for all four tissues, with the largest gains observed for VAT (mean single-region R2 0.84 vs combined 0.98) and SAT (0.88 vs 0.92).

Prediction of Whole-Body Tissue Composition from Regional Sub-Body CT Scans

期刊论文
发表日期：

2026年7月21日

查看全文

Morteza Golzan,
Hyunwoo Lee,
Vincent Chow,
Telex M. N. Ngatched,
Lihong Zhang,
Da Ma,
Maciej Michalak,
Karteek Popuri,
Mirza Faisal Beg

摘要

Accurate assessment of body composition is essential for understanding human physiology and health risks and developing personalized medical strategies. Traditional approaches, such as single-slice segmentation at the L3 vertebra, often fail to capture the complexities of whole-body tissue distribution, which is influenced by genetics, metabolism, environment, and physiology. To address these limitations, we present a novel methodology to estimate whole-body composition using sub-body anatomical region computed tomography (CT) scans. For this study, a cohort of 101 subjects with plasmacytoma cancer who underwent whole-body (head-to-toe) CT scans was utilized. Using in-house segmentation software, skeletal muscle (SKM), subcutaneous adipose tissue (SAT), visceral adipose tissue (VAT), and intermuscular adipose tissue (IMAT) were quantified. Generally, sub-bodies were either individual regions, such as the chest (CHE), abdomen (ABD), and pelvis (PLV), or combined regions, including chest and abdomen (CHA) and abdomen and pelvis (ABP), as defined according to standard vertebral landmarks. A multivariate linear regression model incorporating demographic and CT-derived features is developed to predict whole-body tissue volumes. Tenfold cross-validation was used to validate the performance of the prediction models, and evaluation metrics included the mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), 95% confidence intervals across folds, and the coefficient of determination (R²). For individual regions, SKM demonstrates strong performance with R² values from 0.83 to 0.91 and low MAE (e.g., chest 1.14 ± 0.83 L, R² = 0.913). SAT performs best in the PLV region (R² = 0.918, MAE = 1.58 ± 1.28 L), while VAT achieves its highest single-region performance in the ABD region (R² = 0.950, MAE = 0.37 ± 0.35 L). IMAT shows small MAE values (e.g., chest 0.26 ± 0.21 L) but relatively low R², with the highest predictability obtained in the CHE region (R² = 0.822). For combined regions, SKM performs best in the CHA region (R² = 0.924, MAE = 1.01 ± 0.84 L), while SAT improves in the ABP region (R² = 0.923, MAE = 1.45 ± 1.35 L). VAT shows consistently excellent performance across both combined regions, with similar R² values (CHA 0.977, ABP 0.981) and low MAE. Broader anatomical coverage improved predictions for all four tissues, with the largest gains observed for VAT (mean single-region R² 0.84 vs combined 0.98) and SAT (0.88 vs 0.92).