Nonrigid registration method for longitudinal chest CT images in COVID-19

Abstract

Rationale and objectives: To analyze morphological changes in patients with COVID-19-associated pneumonia over time, a nonrigid registration technique is required that reduces differences in respiratory phase and imaging position and does not excessively deform the lesion region. A nonrigid registration method using deep learning was applied for lung field alignment, and its practicality was verified through quantitative evaluation, such as image similarity of whole lung region and image similarity of lesion region, as well as visual evaluation by a physician.

Materials and methods: First, the lung field positions and sizes of the first and second CT images were roughly matched using a classical registration method based on iterative calculations as a preprocessing step. Then, voxel-by-voxel transformation was performed using VoxelMorph, a nonrigid deep learning registration method. As an objective evaluation, the similarity of the images was calculated. To evaluate the invariance of image features in the lesion site, primary statistics and 3D shape features were calculated and statistically analyzed. Furthermore, as a subjective evaluation, the similarity of images and whether nonrigid transformation caused unnatural changes in the shape and size of the lesion region were visually evaluated by a pulmonologist.

Results: The proposed method was applied to 509 patient data points with high image similarity. The variances in histogram characteristics before and after image deformation were confirmed. Visual evaluation confirmed the agreement between the shape and internal structure of the lung field and the natural deformation of the lesion region.

Conclusion: The developed nonrigid registration method was shown to be effective for quantitative time series analysis of the lungs.

Keywords: Chest computed tomography (CT); Coronavirus disease 2019 (COVID-19); Deep learning; Nonrigid registration.

Fig. 1

Flowchart of the patient population. Note: COVID-19 = coronavirus disease 2019. Of the 819 participants, 81 were excluded because of the following reasons: age <20 years (n = 31), pregnancy (n = 3), no CT scans (n = 32), hospital transfer (n = 1), and data mismatch (n = 14). Another 229 patients were excluded because they had not undergone longitudinal scans.

Fig. 2

Network architecture for VoxelMorph.

Fig. 3

Objective evaluation scheme. Note: First CT images (a) and second CT images (c) and (d) were compared.The second CT images were generated from the original image (b), an image with affine transformation (c), and an image with affine transformation + VoxelMorph (d).Comparisons were made with the first CT images.

Fig. 4

The same-level slices for the upper end of the arch (Axial 1), trachea (Axial 2), and lower pulmonary vein inlet (Axial 3) are shown ((a)–(b)). The tracheal branch (coronal) and right lung center (sagittal) of the aorta in the first CT image are shown for each image ((b)–(d)). Note: First CT image (a); second CT image (b); affine alignment result of second CT image (c); VoxelMorph applied to second CT image (d).

Fig. 5

Image similarity evaluation results between the first CT image and affine alignment or affine + VoxelMorph results. Comparison of zero mean normalized cross-correlation (a), comparison of structural similarity (b), comparison of the peak signal-to-noise ratio (c), and dice coefficient of pulmonary vascular region (d).

Fig. 6

P value summary of the Wilcoxon signed-rank test for first order statistics and 3D shape features before and after nonrigid registration in the lung lesion region.

Fig. 7

Summary of visual evaluation results in the three axial level slices. Similarity of the structure (a). Similarity of lesion region shape, size, and texture of COVID-19 pneumonia (b).

Fig. 8

Example of low-similarity results. The upper end of the arch (Axial 1), tracheal branch (Axial 2), lower pulmonary vein inlet (Axial 3), tracheal branch (coronal), and right lung center (sagittal) of the aorta in the first CT image are presented for each image. First CT image (a); second CT image (b); affine alignment result of second CT image (c); VoxelMorph application to second CT image (d).