Fig. 1.

Flowchart of the proposed hybrid label learning. Step 1 trains the infected region segmentation network UNet-1 using fully supervised learning. UNet-1 also provides initialization for the consolidation segmentation network UNet-2. Step 2 trains UNet-2 combining the image-level label and prior probability built from images . The solid brown lines are the procedures that are involved in both training and testing, whereas the dashed blue lines are training procedures only.