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

Since the outbreak of COVID-19, there have been increasing efforts on developing deep learning methods to perform screening of COVID-19 based on medical images such as CT scans and chest X-rays. Wu [6] . Several works have also applied 3D deep learning models to screen COVID-19 based on chest CT images [7] , [8] . Yang et al. developed a deep learning based CT diagnosis system (DeepPneumonia) to assist clinicians to identify patients with COVID-19 [9] . Xu et al. developed a deep learning algorithm by modifying the inception transfer-learning model to provide clinical diagnosis ahead of the pathogenic test [10] . Shi et al. employed the "VB-Net" neural network to segment COVID-19 infection regions in CT scans [11] . Yu et al. constructed a system based on UNet++ for identification of COVID-19 from CT images [12] . Shen et al. proposed an infection-size-aware Random Forest (iSARF) method which can automatically categorize subjects into groups with different ranges of infected lesion sizes [13] .