Classification of X-ray images into COVID-19, pneumonia, and TB using cGAN and fine-tuned deep transfer learning models

Mehta, Tirth; Mehendale, Ninad

PURPOSE: The rapid increase in the spread of the coronavirus disease of 2019 (COVID-19) has led to a need for reliable, effective, and readily available testing on a large scale. While diagnostic testing has been a support to public health, newer technology can be used to provide low-cost and convenient test options for patients. X-ray scanning can be performed to resolve this issue and produce quicker and more precise results. Currently, a radiologist is required to examine these X-ray images. However, deep convolutional neural networks can also be used to perform X-ray examinations and employed for the detection of COVID-19. We propose a Conditional Generative Adversarial Network (cGAN) with a fine-tuned deep transfer learning model to classify chest X-rays into six categories: COVID-Mild, COVID-Medium, COVID-Severe, Normal, Pneumonia, and Tuberculosis. METHODS: A total of 1229 images were taken to form a dataset containing six classes corresponding to the six categories. A cGAN was used to increase the number of images. Generative Adversarial Networks (GAN) are used to train a model for generating new images. cGAN is an extension of GAN consisting of a generator and discriminator network that are trained simultaneously to optimize the model. The generated images were then trained using deep transfer learning models such as ResNet50, Xception, and DenseNet-169 to achieve the classification into six classes. RESULTS: The proposed model helped achieve a training and validation accuracy of up to 98.20 % and 94.21 % respectively. The model was able to achieve a test accuracy of 93.67 %. The use of cGAN not only helped to increase the size of the training dataset but it also helped to reduce the problem of over-fitting. CONCLUSION: The proposed approach will help to diagnose COVID-19 quickly at an early stage. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s42600-021-00174-z.