Data augmentation on chest X-rays for improving pathology classification performance

In recent years, deep learning techniques have made great progress. We can see applications of it in many fields such as economics, military, healthcare, and so on. Healthcare, in particular, is one of the most critical of these areas. While the world population is growing every day, healthcare professionals need more computerized technologies to make things faster. Proposed new methods are making important contributions to the healthcare system, but the lack of data is limiting development. Privacy issues prevent more patient data from being collected to use for training models. For example, chest X-rays are commonly used in pathology classification. However, studies are limited due to the lack of public datasets. To solve this problem, we focus on data augmentation on chest X-rays to improve pathology classification results. To this end, we demonstrate three methods. In the first, we propose a heatmap based image inpainting that uses X-ray images with observations and inpaints the large healthy areas to create new X-rays while preserving the labels. The second proposed method synthesizes images using an extended version of GANSpace by adding a conditional generator StyleGAN2-ADA. Finally, we demonstrate the manipulation of real and healthy X-ray images using latent space manipulation and GAN inversion. Our quantitative experiments show that heatmap based inpainting improves classification results from 86.1\% to 87.7\%. To provide a basis for our Conditional GANSpace method, the results of X-ray image generation experiments using StyleGAN2-ADA are also provided. The classification result of the dataset augmented using StyleGAN2-ADA is 87.36\% and our Conditional GANSpace improves this result with the highest result of 88.5\%.