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

In this work, we aim to address these two problems by (1) building a publicly-available dataset containing hundreds of CT scans that are positive for COVID-19 and (2) developing sample-efficient deep learning methods that can achieve high diagnosis accuracy of COVID-19 from CT scans even when the number of training CT images are limited. We first collect the COVID19-CT dataset, which contains 349 CT images with clinical findings of 216 COVID-19 patient cases. The images are collected from medRxiv and bioRxiv papers about COVID-19. CTs containing COVID-19 abnormalities are selected by reading the figure captions in the papers. We manually remove artifacts in the original images, such as texts, numbers, arrows, etc. Figure 1 shows some examples of the COVID-19 CT scans. To our best knowledge, it is the largest COVID-19 CT dataset to date. And all the images are open to the public for research purpose. Given this dataset, we develop deep learning (DL) methods to perform CT-based diagnosis of COVID-19. Though largest among its kind, COVID19-CT is still limited in image number. DL models are data-hungry, which have high risk of overfitting when trained on smallsized dataset. To address this problem, we develop sampleefficient methods to train highly-performant DL model in spite of data deficiency. Specifically, we investigate two paradigms of learning approaches for mitigating data deficiency: transfer learning and self-supervised learning.