Detection of Pneumonia in Chest X-Ray Images Using Deep Transfer Learning and Data Augmentation With Auxiliary Classifier Generative Adversarial Network
C. Cala-or, A. Ambita, A. Carajay, and J. Sanz (34-54)
Abstract
Deep learning applications in medical research are often constrained by the lack of data availability due to the significant labor and cost required to collect data. Such issues cause the convolutional neural networks (CNNs) to suffer with overfitting and a drastic loss in accuracy. To overcome this problem, generative adversarial networks (GANs) have been adopted in medical imaging as a data augmentation technique because of their capability to generate realistic samples that help add variability in the training set. Therefore, this paper proposes a data augmentation based on GAN to overcome the issue of limited data availability in conjunction with pretrained CNN models on detecting pneumonia from chest x-ray images. We use auxiliary classifier GAN (ACGAN), which extends traditional GAN by making the generation of images conditional on a side information such as labels. The proposed method has further improved the performance of the CNN models most especially the ResNet variants that improved by more than 10%. ResNet-18, the smallest ResNet variant, showed the highest improvement with 13.36% in accuracy and 16.13% in F1-score and also outperformed the other CNN models used in the experiment. The addition of ACGAN-generated images has proven to be effective in adding variability to the training set.