Agreement of automated computed tomography body composition measurement across image reconstruction algorithms under identical projection data

CT-based body composition analysis provides quantitative biomarkers of muscle, adipose tissue, and bone mineral density. However, whether reconstruction techniques influence automated measurements when radiation exposure is held constant remains unclear. This study aimed to determine the effect of reconstruction on CT-based body composition and bone mineral density (BMD) using fully automated deep learning segmentation. This retrospective study included 202 adult patients (mean age 65.5 years, n = 112 men) who underwent abdominopelvic CT between April and December 2025. All scans were reconstructed from identical raw projection data using filtered back projection (FBP), hybrid iterative reconstruction (IR), and deep learning reconstruction (DLR). Body composition was quantified at the third lumbar vertebral level (L3), and BMD was assessed at the first and second lumbar vertebral bodies (L1 and L2) using an automated deep learning–based segmentation tool. Quantitative parameters included adipose tissue density and area, muscle density and area, organ attenuation and volume, and BMD. Agreement across reconstruction algorithms was assessed using Bland–Altman analysis with 95% limits of agreement (LOA) and intraclass correlation coefficients (ICCs). Noise characteristics were assessed using noise power spectrum analysis. Most adipose and muscle area measurements demonstrated small absolute differences (≤ 1–2 cm²), whereas attenuation-threshold muscle-composition parameters, including low-attenuation and normal-attenuation muscle areas, showed greater reconstruction-related variability. Bland–Altman analysis demonstrated low bias and narrow limits of agreement for subcutaneous fat area (bias, 0.05 cm²; 95% LOA, − 1.51 to 1.61 cm²; FBP vs. DLR), skeletal muscle area (bias, − 0.32 cm²; 95% LOA, − 1.42 to 0.76 cm²; FBP vs. hybrid IR), and visceral fat area (bias, − 0.56 cm²; 95% LOA, − 2.04 to 0.91 cm²; FBP vs. hybrid IR). BMD showed a bias of 0.12 mg/cm³ (95% LOA, − 1.93 to 2.18 mg/cm³; hybrid IR vs. DLR). ICCs ranged from 0.953 to 1.000 across the evaluated measurements, with the lowest ICC observed for low attenuation muscle area between FBP and DLR. Noise magnitude progressively reduced from FBP to hybrid IR to DLR. Under the evaluated single-scanner protocol, conventional muscle and adipose tissue area measurements showed relatively small reconstruction-related differences. However, attenuation-threshold muscle-composition measurements, particularly low- and normal-attenuation muscle areas, differed substantially across reconstruction settings. These findings support the use of consistent reconstruction settings when attenuation-dependent biomarkers are used and require validation across scanners, protocols, software platforms, and longitudinal examinations.