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

Transfer learning is normally performed by taking a standard neural architecture along with its pretrained weights on large-scale datasets such as ImageNet [20] , and then finetuning the weights on the target task. This idea has been successfully applied to visual recognition [21] as well as language comprehension [22] . In the medical domain, transfer learning has also been widely used in medical image classification and recognition tasks, such as tumor classification [23] , retinal diseases diagnosis [24] , pneumonia detection [25] , and skin lesion and cancer classification [26] , [27] . A recent study in [28] explores the properties of transfer learning for medical imaging tasks and finds that the standard large networks pretrained on ImageNet are often over-parameterized and may not be the optimal solution for medical image diagnosis. In this paper, we continue to investigate different strategies of transfer learning and integrate contrastive self-supervised learning into the transfer learning process to learn powerful and unbiased feature representations for reducing the risk of overfitting.