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

In this section, we evaluate the performance of our proposed Self-Trans networks, and compare them with networks pretrained on large-scale datasets. Given the weights pretrained on other datasets, we leverage contrastive self-supervised learning (CSSL) to further train these weights on CT images in our COVID19-CT dataset. Note that in this step, the labels of these CT images are not utilized. CSSL is only performed on the CTs themselves. After CSSL training, we fine-tune the weights on both the CTs and the class labels by optimizing the classification loss. 1) Self-supervised baselines: To address the effectiveness of self-supervision, we also establish a baseline model with self-supervised auxiliary task. In this paper, we select the image rotation prediction [32] as auxiliary task in multi-task learning scheme. For each training image x, we randomly rotate it with angle φ ∈ {0 • , 90 • , 180 • , 270 • }. A 4-way rotation prediction classifier and CT classification classifier share the same feature extractor. Losses for both tasks are added together and model is jointly trained. We do not rotate samples at test time.