Sample-Efficient Deep Learning for COVID-19 Diagnosis Based on CT Scans

Xuehai He; Xingyi Yang; Shanghang Zhang; Jinyu Zhao; Yichen Zhang; Eric Xing; Pengtao Xie

• To train and evaluate the system, we collect the COVID19-CT dataset 2 , which contains 349 positive CT scans with clinical findings of COVID-19, and 397 negative images without findings of COVID-19. To the best of our knowledge, this is the largest publicly-available CT dataset for COVID-19. • We design different transferring strategies and perform a comprehensive study to investigate the effects of transfer learning for COVID-19 diagnosis and provide insightful findings. • To learn from limited labeled data, we propose Self-Trans networks, which synergistically integrate contrastive selfsupervised learning with transfer learning to learn powerful and unbiased feature representations for reducing the risk of overfitting. • We perform extensive experiments to demonstrate the effectiveness of our proposed methods. It achieves an F1 score of 0.85, an AUC of 0.94, and an accuracy of 0.86 on the COVID19-CT dataset. The rest of the paper is organized as follows. Section 2 reviews related works. Section 3 and 4 present the dataset, methods, and experiments. Section 5 concludes the paper.