COVIDDCGAN: Oversampling Model Using DCGAN Network to Balance a COVID-19 Dataset

Abstract

The COVID-19 infection was announced as a pandemic in late 2019. Due to the high speed of the spread, rapid diagnosis can prevent the virus outbreak. Detection of the virus using prominent information from CT scan images is a fast, cheap, and accessible method. However, these image datasets are imbalanced due to the nature of medical data and the lack of coronavirus images. Consequently, the conventional classification algorithms classify this data unsuitably. Oversampling technique is one of the most well-known methods that try to balance the dataset by increasing the minority class of the data. This paper presents a new oversampling model using an improved deep convolutional generative adversarial network (DCGAN) to produce samples that improve classifier performance. In previous DCGAN structures, the feature extraction took place only in the convolution layer, while in the proposed structure, it is done in both the convolution layer and the pooling layer. A Haar transform layer as the pooling layer tries to extract better features. Evaluation results on two hospital datasets express an accuracy of 95.8 and a loss criterion of 0.5354 for the suggested architecture. Moreover, compared to the standard DCGAN structure, the proposed model has superiority in all classification criteria. Therefore, the new model can assist radiologists in validating the initial screening.