Fig. 2.

An overview of our network architecture for COVID-19 diagnosis (best viewed in colour). Each augmented CT image is fed into the pre-trained periphery-aware encoder, generating representations in $d_e$-dimension. A classifier is learned on top of the representations for pneumonia classification. Meanwhile, the representations are further enhanced in a contrastive learning manner, discriminating the positive (orange arrows) and negative (blue arrows) pairs after being mapped by the projection network into the $d_p$-dimensional space. The enhanced representations can promote more precise diagnostic performance. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)