Fig. 4

Representative predicted and ground truth PRMs of 6 research participants (A–F) in the test set. The left 3 columns (1–3) are predicted PRMs based on single inspiratory chest CT scan using deep learning, and the right 3 columns (4–6) are ground truth PRMs from real inspiratory and expiratory CT scans. In the PRM, red represents emphysema, yellow represents fSAD, and green represents normal areas. Participants A and B have moderate CLE on CT with bronchial dilation and inflammatory SAD. Both predicted and ground truth PRMs show a large range of emphysema and fSAD areas (GT PRM^Emph, Pred PRM^Emph, GT PRM^fSAD, Pred PRM^fSAD for participant A were 7.1%, 8%, 35%, 35.8%, respectively, and for participant B were 8%, 8.9%, 29.5%, 32.9%, respectively). Participant C has only focal CLE on CT with slight bronchiectasis and no inflammatory SAD, but both predicted and ground truth PRMs show a not small fSAD area (GT PRM^fSAD, Pred PRM^fSAD being 27.9%, 35.9%, respectively). Participants D–F have no abnormalities on CT. Participants D and E still have fSAD areas in PRM, but their lesion volume percentages are not high (GT PRM^fSAD, Pred PRM^fSAD for participant D were 11.3%, 17.1%, respectively; GT PRM^fSAD, Pred PRM^fSAD for participant E were 12%, 16.7%, respectively). Participant F has a perfect lung both visually and quantitatively. Overall, the similarity between the emphysema regions in the predicted and ground truth PRMs is high, but the predicted PRM has relatively less fSAD in dependent lung areas and more fSAD in non-dependent lung areas