Sex estimation of the pelvis by deep learning of two-dimensional depth images generated from homologous models of three-dimensional computed tomography images

The utility of convolutional neural networks (CNNs) for sex estimation of the pelvis was evaluated using depth images generated from reconstructed three-dimensional (3D) computed tomography images. The 3D volume data were normalized by a homologous modeling technique to create polygon data with identical topology, then captured images for learning and testing. The neural networks were trained via transfer learning. As a result, a correct assignment rate >90% was obtained in most trials. The frontal view of the pelvis with 60-degree inclination achieved the best results. Selecting samples close to the average images of the sex was effective for training.