Comparação entre segmentação automatizada e manual em tomografias computadorizadas para análise da composição corporal

Abstract

Body composition (BC) assessment using computed tomography (CT) has become an important tool in the investigation of clinical, prognostic, and epidemiological outcomes, particularly through morphometric analysis at the level of the third lumbar vertebra (L3), which is widely recognized as the gold standard for quantifying skeletal muscle and adipose tissue compartments. However, manual segmentation of these structures remains an operator-dependent and time-intensive process, representing a significant barrier to the systematic application of these biomarkers in large-scale studies and Big Data contexts. In this context, the present study aimed to validate the Automated Identification of Body Composition (IACC), a computational tool developed for automated L3 level identification and segmentation of the main body compartments, by comparing its performance with the reference method based on manual segmentation. A retrospective cross-sectional study was conducted, including the analysis of 126 abdominal CT scans obtained from routine clinical examinations. The ground truth was established by a trained evaluator and validated by a senior radiologist, demonstrating high accuracy, as indicated by DICE coefficients greater than 0.93. Statistical analyses included agreement assessment using the intraclass correlation coefficient (ICC), linear correlation (Pearson), spatial overlap (Dice), and systematic bias evaluation through Bland–Altman plots. The results demonstrated high concordance between the automated method and manual segmentation, corresponding to an operational efficiency gain exceeding 80%. Segmentation of adipose tissue compartments showed consistent performance, with particular emphasis on visceral adipose tissue (VAT), which exhibited minimal bias (−0.35 cm²) and an ICC of 0.958. For skeletal muscle, a strong correlation with the reference method was observed; however, Bland–Altman analysis revealed the presence of systematic bias,but not significant, characterized by a mean underestimation of −12.79 cm², a limitation commonly reported in the literature for automated approaches. In conclusion, the IACC represents an effective, reproducible, and computationally efficient alternative to manual body composition segmentation in CT.