Fig. 2

The proposed pipeline for lung segmentation and densitometry. First, a multiclass network is used to identify the lungs, heart, and airways. Heart and airways segmentations are both needed to compute transfer function from raw grey levels into Hounsfield units (HU). Lung segmentation is used to identify the region of interest for the subsequent analysis. Then, three different higher-resolution 2D networks are used to segment the right and left lung on the three views (axial, sagittal, and coronal) of the HU-converted and cropped µCT scans. The result of these three networks is integrated to obtain an accurate and spatially coherent segmentation and thresholded in the [− 1040, + 121] HU range to remove voxels labeled by the network as lungs but whose corresponding value in the µCT scan is outside the appropriate range. Finally, this clean segmentation is compartmentalized based on the corresponding voxel value in µCT according to the thresholds introduced by [20]