Medical Data Augmentation Using Generative Adversarial Networks: X-ray Image Generation for Transfer Learning of Hip Fracture Detection

In the medical domain, it is difficult to retrieve large datasets. Classic data augmentation methods such as flipping, rotating, and scaling, are helpful in classification tasks, but these methods usually are not good enough to increase diversities and variances in small datasets. In this study, we applied AC-GAN (Auxiliary Classifier GANs) for data augmentation of our Limb X-ray dataset. We pre-trained the model for hip fracture detection of our Pelvic X-ray dataset by utilizing transfer learning with the augmented Limb X-ray dataset, which contained both the original dataset and realistic synthetic images made by the AC-GAN. The final hip fracture classification results of the Pelvic X-ray dataset showed that our generative model not only succeeded in producing realistic Limb X-ray data but also helped improve the performance of the transfer learning model for hip fracture detection.