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 aims to leverage data-rich source-tasks to help with the learning of a data-deficient target task (CT-based diagnosis of COVID-19 in our case). One commonly used strategy is to learn a powerful visual feature extraction deep network by pretraining this network on large datasets in the source tasks and then adapt this pretrained network to the target task by finetuning the network weights on the smallsized dataset in the target task. While effective in general, transfer learning may be suboptimal due to the fact that the source data may have a large discrepancy with the target data in terms of visual appearance of images and class labels, which causes the feature extraction network biased to the source data and generalizes less well on the target data. We design different transferring strategies and perform a comprehensive study in the dimensions of source-target domain difference and neural architectures to investigate the effects of transfer learning for COVID-19 diagnosis and provide insightful findings.