Keywords:
Artificial Intelligence, CT, Computer Applications-General, Outcomes
Authors:
J. Pan, K.-H. Nenning, S. Röhrich, N. Sverzellati, V. Poletti, J. Hofmanninger, A. Makropoulos, H. Prosch, G. Langs
DOI:
10.26044/ecr2023/C-14403
Methods and materials
The study enrolled 55 patients who had been diagnosed with IPF, and each patient underwent at least two subsequent CT scans. Using an image analysis deep learning algorithm, the lung texture patterns in each scan were automatically segmented, and lung voxel was assigned to one of six distinct disease patterns based on their appearance in the CT scan. These patterns included ground-glass opacity, reticular pattern, emphysema, nodular pattern, honeycombing, and consolidation. The volume of the total voxels containing these six patterns was calculated for each scan. The overall ratio of the distinct lung texture volume to the total volume was used as a static pattern profile. Thus, each patient had two static pattern profiles based on their two CT scans. In addition, the change in the volume of each pattern over time was evaluated as a dynamic pattern profile, which was calculated based on the difference between subsequent CT scans. Therefore, each patient also had a dynamic pattern profile reflecting the changes in the patterns over time. To evaluate the association of each distinct CT pattern and its change with patient outcomes, we conducted Cox proportional hazard regression analysis for each pattern individually. This allowed us to test the potential impact of each CT pattern on patient survival.