Generation of Synthetic X-Rays Images of Rib Fractures Using a 2D Enhanced Alpha-GAN for Data Augmentation

X-rays are the most commonly performed medical imaging tests to detect fractures. However, some fractures are difficult to detect and may go unnoticed by physicians. In addition, there are very few public X-rays datasets of rib fractures. Although, the creation of such datasets is very time-consuming because of the bureaucratic and ethical issues involved, it is very useful, because these images can be used for teaching and data enhancement without privacy issues. Recently, generative models have been used to synthesize images with high quality and realism. In this work, a generative model was developed to generate synthetic X-ray images with hard-to-detect rib fractures. These images were evaluated using quantitative metrics and a Turing test. It was found that the images generated were not realistic enough due to the large heterogeneity of the dataset used for training, which made it impossible for the model to correctly evaluate the most important features.