Applying general adversarial networks in convolutional neural networks of the 2D whole os coxae image classification for sex estimation in a Thai population

Sex estimation is imperative in forensic identification. Our study aimed to apply a convolutional neural network (CNN) for classifying 2D whole os coxa images and to use generative adversarial networks (GANs) for improving a small dataset for greater accuracy in sex estimation. This study will develop a reliable and accurate technique for adult sex estimation. The samples consisted of 250 left os coxal bones. The age and stature at death ranged from 26 to 94 years, and 135 to 190 cm, respectively. In this study, we used GoogLeNet for classifying sex based on the 2D os coxae images and applied the GANs technique for generating images to solve the small dataset problem. The results showed the validation accuracies were 93.33% and 97.78% of the original sample only and the combined original with GAN samples, respectively. The test accuracies showed 92% of the original sample only, and 96% of the combined original with GAN samples. In conclusion, 2D whole os coxae images performed exceptionally well for classifying sexes using CNNs in Thai samples, and GANs can improve a small dataset effectively. In a further study, this method could derive a large number of the training dataset for improving CNN model performance.