Comparative Evaluation of Deformable Image Registration (DIR) Algorithms
Abstract
Purpose
To compare the registration accuracy of three deformable image registration (DIR) algorithms—DIR option in Radformation AutoContour, Velocity Single-Pass (SP), and Velocity Multi-Pass (MP)—using both phantom and patient datasets.
Methods
Phantom Evaluation: Multiple configurations of molded phantoms (Play-Doh and charcoal) were scanned, including rotations and twisting to introduce known geometric deformations. Deformed configurations were registered back to their corresponding non-deformed reference images. Patient Evaluation: Ten head-and-neck CT datasets were analyzed. In the absence of longitudinal scans, anatomically similar patient pairs were identified, and reference CTs were deformed using Velocity-based workflows to generate realistic secondary datasets, which were then registered back to the original images. Registration accuracy for different algorithms was evaluated by comparing Dice Similarity Coefficient (DSC), 95th percentile Hausdorff Distance (HD95), and Mean Surface Distance (MSD) for selected structures on registered and original images using in-house Python/SimpleITK software.
Results
Phantom Data: All three systems achieved high volumetric agreement (mean Dice ≥ 0.90). Velocity Multi-Pass demonstrated the most consistent performance, exhibiting the narrowest metric range (Dice = 0.94 ± 0.04; HD95 = 3.07 mm; MSD = 1.02 mm), while Radformation AutoContour and Velocity Single-Pass produced comparable mean values with greater variability for complex geometries. Patient Data: Velocity Multi-Pass and Radformation AutoContour demonstrated clinically comparable performance, maintaining sub-millimeter and low-millimeter Mean surface distances, respectively, for most organs at risk. Velocity Single-Pass showed reduced accuracy in several soft-tissue structures, with Dice values as low as 0.67 for the submandibular gland. High-contrast structures (mandible, brainstem) consistently showed the highest agreement, whereas smaller, low-contrast structures (optic nerves, parotids) exhibited greater variability across all algorithms.
Conclusion
Both Radformation Auto Contour and Velocity Multi-Pass provide high and clinically comparable registration accuracy, significantly exceeding the Velocity Single-Pass performance in complex soft-tissue regions.