Supervised InfNet

The supervised InfNet (SInfNet) is a recently developed CNN model for COVID-19 lung CT segmentation [22], which was used as both our backbone and one of the baseline models. We did not change the overall structure and default hyperparameters of the original SInfNet model (Additional file 1: Figure S1). A complete SInfNet consists of two parts: a single SInfNet (Additional file 1: Figure 2A) and a multi SInfNet (Additional file 1: Figure 3A). The single SInfNet only predicts the infected region without classifying them more specifically. The input of the single SInfNet is a raw CT lung image and the output includes the edge contour of the overall lesion regions and four overall lesion region segmentations with different sizes as shown in Additional file 1: Figure S1. A CT lung image is first passed into the initial convolutional layers of the single InfNet to extract image features. Then, the features generated from the convolutional layer are fed into the partial decoder module, reverse attention module, and the edge detection module. The edge detection module is meant to help the network with the detection of the boundaries of the segmentation. The reverse attention and the partial decoder generate the segmentation of the infection regions of the CT lung images.

The prediction from the single SInfNet represents the overall infected regions and acts as a prior to be fed, concatenated with the original CT images, into the multi SInfNet. The multi SInfNet is used to predict multiple labeled segmentations. The segmentations include the predicted background, GGO, and consolidation.