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 mitigate the inefficiency and shortage of existing tests for COVID-19, many efforts have been devoted to searching for alternative testing methods. Several studies [1] have shown that computed tomography (CT) scans manifest clear radiological findings of COVID-19 patients and are promising in serving as a more efficient and accessible testing manner due to the wide availability of CT devices that can generate results at a fast speed. Further, to alleviate the burden of medical professionals in reading CT scans, several works [2] have developed deep learning methods that can automatically interpret CT images and predict whether the CTs are positive for COVID-19. While these works have shown promising results, they have two limitations. First, the CT scan datasets used in these works are not sharable to the public due to privacy concerns. Consequently, their results cannot be reproduced and the trained models cannot be used in other hospitals. Besides, the lack of opensourced annotated COVID-19 CT dataset greatly hinders the research and development of more advanced AI methods for more accurate CT-based testing of COVID-19. Second, these works require a large collection of CTs during model training to achieve performance that meets the clinical standard. Such a requirement is stringent in practice and may not be met by many hospitals, especially under the circumstances that medical professionals are highly occupied by taking care of COVID-19 patients and are unlikely to have time to collect and annotate a large number of COVID-19 CT scans.