Fig. 5

Representative segmentations from the U-Net algorithm from four slices in four healthy subjects. Red shows expected output mask, green shows model output mask, and orange shows the overlap between the two masks. The U-Net created segmentation masks are comparable to the expected output (volunteers 1, 2) and successfully identified image slices where no lung tissue was present (volunteer 3). However, the model mistakenly identified the lower airways as lung in 0.37% of the annotated pixels (volunteer 4)