X-Ray Images Dataset Augmentation with Progressively Growing Generative Adversarial Network

Abstract

This paper proposes a deep learning based architecture to mitigate the problem of imbalance medical dataset due to non-availability of adequate data and privacy policy of medical organizations. A modified Progressively Growing Generative Adversarial Network (PGGAN) is used to create good quality of artificial X-ray images to augment the original dataset. A dataset having 2,000 X-ray images per class representing four different pathologies is used by the modified PGGAN for data augmentation. Transfer learning is used to adopt standard classification models which are fine tuned to suit this application. Fine tuning is done by re-training the classification models with the newly created augmented and balanced dataset (original X-ray images and artificially created X-ray images). The augmented dataset containing a total of 56,000 X-ray images are used to train and 4,000 original images are used to test the classification models. The experimental results show that the use of modified PGGAN for data augmentation and transfer learning for classification models increases the accuracy of predicting Chest pathologies.