Self-supervised deep learning for improving efficiency of COVID-19 screening on chest X-ray

Abstract

As we face the global pandemic of COVID-19, there is an increasing effort to use deep learning model to support chest X-ray COVID-19 diagnosis. However, the lack of COVID-19 chest X-ray dataset is a major roadblock in training the model, resulting in low accuracy of models. To solve the data shortage problem, methods like transfer learning using images from other domain and generate artificial training data using generative adversarial network (GAN) have been proposed, limitations in reliability due to domain differences and using artifacts are apparent. We propose to improve the data efficiency and classification accuracy of a COVID-19 CNN classification through self-supervised learning approach, which uses existing data only without adding data to pretrain the model. In this study, we introduce the patch location prediction method, inspired by imitating the chest X-ray reading process for identifying organ positions, which aims to improve the understanding of the X-ray image by allowing the model to learn the structure of the human anatomy. Experimental results show that our method achieves improved performance than previous methods without using irrelevant or artificially generated data for training the model, improving accuracy and reliability.