Estimativa de sexo em brasileiros utilizando mensurações orbitárias em tomografias computadorizadas : abordagens convencional e por aprendizado de máquina

Abstract

Human identification requires resources that match the complexity of the expertise involved. Estimating sex is essential for establishing an individual's biological profile in the process of identifying human remains. However, the use of foreign parameters in Brazilian samples is often ineffective. This study aimed to evaluate the applicability of computed tomography (CT) in measuring the Orbital Index for sex estimation in individuals from the Brazilian population, comparing multiplanar (MPR) and three-dimensional (3D) reconstructions. The sample was obtained by sample calculation (95% confidence, 5% error) and consisted of 386 CT (CBCT and MSCT) scans of patients performed in private clinics, balanced by sex (193 male and 193 female) and distributed proportionally among the five macro-regions of Brazil. Skull or face CT scans with FOV that allowed full visualization of the orbital anatomical points were included; cases with morphofunctional alterations, surgeries, fractures, neoplasms, and individuals with bone immaturity (<18 years) were excluded. Height, width, and IO measurements were performed in MPR and 3D using the free RadiAnt DICOM Viewer 2025.2 software. Intra- and inter-examiner agreement was assessed in 10% of the sample using the Intraclass Correlation Coefficient (ICC), presenting excellent values (ICC > 0.90), with a standardized protocol and triple-blind masking. Descriptive and inferential statistical analyses were performed using Jamovi 2.3 and Python 3.13.5 software. The normality (Kolmogorov-Smirnov, p > 0.05) and homogeneity (Levene, p > 0.05) of the data were confirmed, and parametric tests were then applied. There was a significant difference between MPR and 3D (paired t, p < 0.05), with a mean IO of 110.59 (± 6.112) in MPR and 100.57 (± 6.451) in 3D, indicating an underestimation of 3D in relation to MPR (≈8.84%; ± 5.01%). Men had larger linear orbital dimensions than women in both reconstructions (p < 0.001), but IO dimorphism was confirmed only in 3D (p < 0.001), following this reconstruction for subsequent analyses. The 3D reconstruction revealed a moderate correlation between the right and left orbital indices and a strong correlation between these measurements and the mean orbital index. Linear discriminant analysis (LDA) conducted with the mean index demonstrated low discriminatory capacity, explaining 5% of the variability between sexes and achieving an overall accuracy of 60.6%. The ROC curve confirmed statistically significant but limited discrimination (AUC = 0.637). Machine learning models (Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree, Logistic Regression, Multi-Layer Perceptron (MLP), Gradient Boosting, and Random Forest) trained with 3D orbital measurements showed accuracies between 0.603 and 0.679 and AUC between 0.590 and 0.700, indicating moderate and relatively homogeneous performance among the algorithms. In all models, orbital widths—especially the right one—were the most relevant variables, while orbital heights showed minimal contribution. The findings confirm that both linear measurements and orbital indices derived from 3D reconstruction have limited discriminatory capacity for sex estimation, even with the use of advanced multivariate methods. The influence of reconstruction (MPR and 3D) highlights the need for specific morphometric parameters for the Brazilian population, ensuring more accurate interpretations in forensic and clinical applications. The findings show that orbital morphology obtained by computed tomography (MPR and 3D) has limited discriminatory capacity for sex estimation, given the low effect of the measurements, the wide overlap between groups, and the moderate performance of statistical and computational approaches. Thus, the orbital index should not be used either as an aid in multivariate models or as isolated indicators.