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

The lack of annotated CT scans about COVID-19 brings significant challenges for deep-learning-based diagnosis of COVID-19 using CT images. To address this problem, we build COVID19-CT, a dataset containing hundreds of CT images positive for COVID-19. Though the largest of its kind, COVID19-CT is still small, making deep learning models trained on this dataset prone to overfitting. To address this problem, we systematically investigate different transfer learning strategies and propose a new approach called Self-Trans, which synergistically integrates unsupervised in-domain selfsupervised learning with supervised out-of-domain transfer learning to learn effective and unbiased visual feature representations that are robust to overfitting. In the following subsections, we first introduce the COVID19-CT dataset that we built, then present transfer learning and the proposed Self-Trans approach for sample-efficient diagnosis of COVID-19 from CT scans.