Abstract
Abstract
Background
Deformable image registration (DIR) is essential for thoracic four‐dimensional computed tomography (4D‐CT)‐based radiotherapy applications. Recently, deep learning‐based DIR methods such as VoxelMorph have been proposed; however, their performance relative to clinically used DIR algorithms remains unclear.
Purpose
This study aimed to evaluate the DIR accuracy of VoxelMorph for thoracic 4D‐CT and to compare it with conventional clinical and research‐oriented DIR methods.
Materials and methods
Thoracic 4D‐CT data from 64 lung cancer patients were retrospectively analyzed. End‐inhalation and end‐exhalation phase images were used for DIR. VoxelMorph was trained using 50 cases, with 4 for validation and 10 for testing. DIR performance on the test dataset was compared with Demons (SimpleITK), modified Demons (Eclipse), and ANACONDA (RayStation). Accuracy was evaluated by mean absolute error (MAE) of CT values computed within the body region, whereas Dice similarity coefficient (DSC) and 95th percentile of Hausdorff distance (HD95) were evaluated within the lung label.
Results
The median MAE decreased from 67.56 HU before DIR to 52.32 HU with modified Demons, 37.39 HU with Demons, 36.18 HU with ANACONDA, and 36.83 HU with VoxelMorph. The median DSC increased from 0.91 to 0.95 for modified Demons, 0.97 for Demons and ANACONDA, and 0.98 for VoxelMorph. The median HD95 was 3.0 mm for modified Demons, 3.1 mm for Demons, 2.0 mm for ANACONDA, and 2.5 mm for VoxelMorph. Overall, VoxelMorph demonstrated competitive accuracy, significantly outperforming modified Demons (adjusted
p
0.05), while showing smaller inter‐case variability and markedly reduced processing time.
Conclusions
VoxelMorph demonstrated DIR performance comparable to that of clinically used algorithms for thoracic 4D‐CT, with high overlap accuracy, relatively low inter‐case variability and shorter processing times under the evaluated implementation conditions. These findings suggest its potential as research‐oriented DIR framework, although further validation under standardized conditions is required before routine clinical implementation.