Three-Dimensional Morphometric Evaluation of the Orbital Aperture in Multislice Computed Tomography: Anatomical Classification and Deep Learning-Based Sex Estimation

Abstract The orbital aperture (OA) is an important anatomical structure that forms the entrance of the orbit and has connections with intracranial structures. This structure has a critical importance for clinicians as it contains reference points in surgical approaches and requires attention in plastic reconstructive surgery. The OA is also one of the craniofacial variables used for sex estimation in anthropology and forensic medicine. The aim of this study was to evaluate the OA morphologically and to determine its role in sex estimation by deep learning method.Three-dimensional (3D) images of 100 adult individuals (48 males, 52 females) on Multislice Computed Tomography (MSCT) were used in the study. A classification for the OA shape was created by classifying the right and left side OA images based on observation by 3 researchers. In this study, sex estimation was performed using convolutional neural network (CNN) models to extract deep features found in OA coronal and MSCT images. Gender prediction percentage was calculated using deep learning method over the OA images registered to the algorithm one by one.For OA morphology, 5 types were identified: square, trapezoid, round, oval, round-trapezoid. The percentage of gender prediction, with the deep learning method, in coronal slices and 3D images was found to be 65.6% and 73.4%, respectively. The most common OA shape was round-trapezoid (74%) in both coronal and 3D images in males, while it was round type (39%) in females.In CNN modeling, incorporating OA types along with side information led to a decrease in gender estimation accuracy. This highlighted the importance of considering the morphological variations of OA and their distribution across sides. Interestingly, when side information is excluded, gender prediction accuracy can exceed 80% in both coronal and 3D images.