Comparison of Affine and DCGAN-based Data Augmentation Techniques for Chest X-Ray Classification

Abstract

Data augmentation, also called implicit regularization, is one of the popular strategies to improve the generalization capability of deep neural networks. It is crucial in situations where there is a scarcity of high-quality ground-truth data. Also getting new samples is expensive and time consuming. This is a typical issue in the medical domain. Therefore, this study compares the performance of Affine and Generative Adversarial Networks (GAN)- based data augmentation techniques on the chest image X-Ray dataset. The Pneumonia dataset contains 5863 chest X-Ray images. The traditional Affine data augmentation technique is applied as a pre-processing technique to various deep learning-based CNN models like VGG16, Inception V3, InceptionResNetV2, DenseNet-169 and DenseNet-202 to compare their performance. On the other hand, DCGAN architecture is applied to the dataset for augmentation. Evaluation measures like accuracy, recall and AUC depict that DCGAN outperforms other traditional models. The most important advantage of DCGAN is that it is able to identify fake images with 100% accuracy. This is especially relevant for the medical domain as it deals with the life of individuals. Thus, it can be concluded that DCGAN has better performance as compared to affine transformations applied to traditional CNN models .