Multi-Center Covid-19 Ct Image-Label Pairs Synthesis Via Few-Shot Generative Adversarial Network Adaptation

Coronavirus disease 2019 (COVID-19) is an ongoing global pandemic that has threatened global health for years. Given its transmissibility and high pathogenicity, an accurate and rapid method of diagnosing COVID-19 infection is essential. Numerous deep learning (DL) models have been developed to assist radiologists with chest computed tomography (CT)-based COVID-19 screening. However, the existing COVID-19 CT datasets usually exhibit significant geographic and class imbalances, resulting in the difficulty of generalizing the model across CT datasets from different patient cohorts. One feasible solution to the problem is to generate data for the small target datasets which have only a few available examples by leveraging a large-scale source dataset as pretraining (i.e., few-shot generative model adaptation). To calibrate the target generative models during adaptation, we present LeCam-CLCR to preserve source images’ diversity information via combining contrastive learning (CL) with LeCam regularization, and make full use of the target images via consistency regularization (CR). We demonstrate the effectiveness of our approach by generating realistic and diverse CT image-label pairs of target datasets, and show that it consistently outperforms the state-of-the-art approaches.