CT image segmentation for inflamed and fibrotic lungs using a multi-resolution convolutional neural network

Abstract

The purpose of this study was to develop a fully-automated segmentation algorithm, robust to various density enhancing lung abnormalities, to facilitate rapid quantitative analysis of computed tomography images. A polymorphic training approach is proposed, in which both specifically labeled left and right lungs of humans with COPD, and nonspecifically labeled lungs of animals with acute lung injury, were incorporated into training a single neural network. The resulting network is intended for predicting left and right lung regions in humans with or without diffuse opacification and consolidation. Performance of the proposed lung segmentation algorithm was extensively evaluated on CT scans of subjects with COPD, confirmed COVID-19, lung cancer, and IPF, despite no labeled training data of the latter three diseases. Lobar segmentations were obtained using the left and right lung segmentation as input to the LobeNet algorithm. Regional lobar analysis was performed using hierarchical clustering to identify radiographic subtypes of COVID-19. The proposed lung segmentation algorithm was quantitatively evaluated using semi-automated and manually-corrected segmentations in 87 COVID-19 CT images, achieving an average symmetric surface distance of [Formula: see text] mm and Dice coefficient of [Formula: see text]. Hierarchical clustering identified four radiographical phenotypes of COVID-19 based on lobar fractions of consolidated and poorly aerated tissue. Lower left and lower right lobes were consistently more afflicted with poor aeration and consolidation. However, the most severe cases demonstrated involvement of all lobes. The polymorphic training approach was able to accurately segment COVID-19 cases with diffuse consolidation without requiring COVID-19 cases for training.

Figure 1

(A) Motivation for polymorphic training. In this work, the desired segmentation target is consolidated cases with specific labels of left lung (LL), right lung (RL), and background (B) (upper right). However, only normal cases with specific labels (upper left) and consolidated cases with non-specific labels of lung (L) and background (B) (lower right) are available for training. The proposed polymorphic training approach allows us to utilize the available training data and generalize to the target domain of consolidated specifically labeled cases (upper right). (B) Standard training (top) using only specifically labeled COPD images lacks the consolidation phenotype necessary to successfully segment injured regions in COVID-19 images. Polymorphic training (bottom) utilizes specifically labeled COPD images with nonspecifically labeled animal models of acute lung injury to achieve specific lung labels including injured regions in COVID-19 images. The specific lung labels are depicted in green and blue for left and right lung, respectively. The nonspecific lung label is depicted in orange.

Figure 2

Polymorphic training accommodates labeled data with different degrees of specificity. In this case some labeled training have specific labels distinguishing left and right lung, while other training data only have a single label for all lung tissue.

Figure 3

Axial slices of CT images (left column) and lung segmentation results for the nonpolymorphic model (center column) and the polymorphic model (right column) algorithms for four COVID-19 patients (by row). Correctly classified voxels are displayed in blue and green for right and left lungs, respectively. False negative and false positive voxels are illustrated in pink and yellow, respectively.

Figure 4

Quantitative evaluation of lung segmentation on the COVID evaluation dataset ($N=87$). The proposed polymorphic model (black) is compared to a nonpolymorphic model (white) using ASSD and the Dice coefficient. Results are stratified by nonaerated lung volume percent in the right panel. Left and right lung results are denoted using left- and right-facing triangles, respectively (left: $◀⊲$, right: $▶⊳$). Linear regression for polymorphic (solid) and nonpolymorphic (dashed) models revealed significantly different coefficients for ASSD in mm ${\%}^{-1}$ (polymorphic: 0.073, nonpolymorphic: 0.138, $p<0.001$) and Dice coefficient in ${\%}^{-1}$ (polymorphic: − 0.003, nonpolymorphic: − 0.006, $p<0.001$).

Figure 5

Sagittal slices of CT images (left column) and right lobe segmentation results for the PTK (center column) and proposed (right column) algorithms for four COVID-19 patients (by row).

Figure 6

Sagittal slices of CT images (left column) and left lobe segmentation results for the PTK (center column) and proposed (right column) algorithms for four COVID-19 patients (by row).

Figure 7

Hierarchical clustering results showing disease subtypes of COVID-19 patients. Each row corresponds to one patient. The left five columns show percent of lobe volume with poor aeration ($-500<\mathrm{HU}<-100$) and the right five columns show percent of lung lobe volume with consolidation ($\mathrm{HU}\ge -100$). Poor aeration is used as an approximation of ground glass opacities. The dendrogram visualization shows four subtypes of patients: (A) mild loss of aeration primarily in the two lower lobes without consolidation, (B) moderate loss of aeration focused in the two lower lobes with or without consolidation in lower lobes, (C) severe loss of aeration throughout all lobes with or without consolidation, and (D) severe loss of aeration and consolidation throughout all lobes.