Sex Prediction From the Clavicle Using Computerized Tomography Images via Traditional and Hybrid Deep Learning Models

The aim of this study is to perform high accuracy sex prediction from clavicle images using proposed hybrid deep learning models and traditional deep learning models. The clavicle of 807 female and 805 male individuals obtained from Computed Tomography were segmented in 3D format and saved in jpeg format as superior-inferior and right-left. MobileNetV2, DenseNet201, and ResNet101 traditional deep learning models and the proposed MobileNetV2+Multilayer Perceptron (MLP) and MobileNetV2+MLP+t-distributed Stochastic Neighborhood Embedding (t-SNE) based hybrid deep learning models were trained with the training set. The training was performed both with and without right-left side discrimination. The performance of each model was evaluated and compared. The highest accuracy rate of 91% was obtained in the training with both proposed hybrid models without side discrimination. The highest success rate obtained with the proposed models was 88%. The lowest accuracy rates were achieved with ResNet101. The accuracy rate was 81% in the analysis with side discrimination and 83% in the analysis without discrimination. According to Grad-Cam, the extremitas sternalis tip contributed the most to accuracy. In this study, MobileNetV2+MLP and MobileNetV2+MLP+t-SNE provided highly accurate results in sex prediction. This approach is a potential new method that can be used in sex estimation, especially in forensic medicine, as it allows the collection of features with MobileNetV2, classification with MLP, visualization with t-SNE, and observation of errors without metric measurement, directly from the superior and inferior available images of the clavicle.