Evaluation of various Deformable Image Registrations for Point and Volume Variations

The accuracy of deformable image registration (DIR) has a significant dosimetric impact in radiation treatment planning. We evaluated accuracy of various DIR algorithms using variations of the deformation point and volume. The reference image (Iref) and volume (Vref) was first generated with virtual deformation QA software (ImSimQA, Oncology System Limited, UK). We deformed Iref with axial movement of deformation point and Vref depending on the types of deformation that are the deformation1 is to increase the Vref (relaxation) and the deformation 2 is to decrease . The deformed image (Idef) and volume (Vdef) acquired by ImSimQA software were inversely deformed to Iref and Vref using DIR algorithms. As a result, we acquired deformed image (Iid) from Idef and volume (Vid) from Vdef. The DIR algorithms were the Horn Schunk optical flow (HS), Iterative Optical Flow (IOF), Modified Demons (MD) and Fast Demons (FD) with the Deformable Image Registration and Adaptive Radiotherapy Toolkit (DIRART) of MATLAB. The image similarity between Iref and Iid was calculated using the metrics that were Normalized Mutual Information (NMI) and Normalized Cross Correlation (NCC). When moving distance of deformation point was 4 mm, the value of NMI was above 1.81 and NCC was above 0.99 in all DIR algorithms.When the Vref increased or decreased about 12%, the difference between Vref and Vid was within 5% regardless of the type of deformation.The value of Dice Similarity Coefficient (DSC) was above 0.95 in deformation1 except for the MD algorithm. In case of deformation 2, that of DSC was above 0.95 in all DIR algorithms. The Idef and Vdef have not been completely restored to Iref and Vref and the accuracy of DIR algorithms was different depending on the degree of deformation. Hence, the performance of DIR algorithms should be verified for the desired applications.

💡 Research Summary

The purpose of this study was to quantitatively assess the performance of several deformable image registration (DIR) algorithms under controlled point‑wise and volumetric deformations, with the ultimate goal of informing clinical use in radiation therapy planning. A reference image (Iref) and a reference volume (Vref) were first generated using the virtual deformation quality‑assurance software ImSimQA (Oncology System Limited, UK). Two deformation scenarios were then created: (1) a localized axial shift of a deformation point (point‑wise deformation) and (2) a global scaling of the entire volume to either increase (relaxation) or decrease (contraction) its size by approximately 12 %. ImSimQA produced the deformed image (Idef) and deformed volume (Vdef) for each scenario.

Four widely used DIR algorithms implemented in the MATLAB‑based Deformable Image Registration and Adaptive Radiotherapy Toolkit (DIRART) were evaluated: Horn‑Schunck optical flow (HS), Iterative optical flow (IOF), Modified Demons (MD), and Fast Demons (FD). Each algorithm was used to inversely deform Idef and Vdef back toward the original reference, yielding a reconstructed image (Iid) and reconstructed volume (Vid).

Image similarity between Iref and Iid was quantified using Normalized Mutual Information (NMI) and Normalized Cross‑Correlation (NCC). Volume similarity was assessed with the Dice Similarity Coefficient (DSC). When the deformation point was displaced by 4 mm, all four algorithms achieved NMI values greater than 1.81 and NCC values above 0.99, indicating excellent global intensity agreement. For volumetric changes of roughly ±12 %, the absolute difference between Vref and Vid remained within 5 % for every algorithm, a tolerance generally acceptable in clinical dose‑distribution calculations.

Algorithm‑specific differences emerged in the DSC analysis. In the relaxation scenario (volume increase), the Modified Demons algorithm produced DSC values below the 0.95 threshold, whereas HS, IOF, and FD all exceeded 0.95. This suggests that MD may over‑smooth the deformation field when the scaling is large, leading to a loss of boundary fidelity. Conversely, in the contraction scenario (volume decrease), all four algorithms maintained DSC values above 0.95, implying that shrinkage deformations are handled more uniformly across methods.

The study highlights several important considerations for clinical implementation of DIR. First, the choice of algorithm should be guided by the expected magnitude and type of deformation: optical‑flow‑based methods (HS, IOF) and fast Demons performed robustly across both point‑wise and volumetric changes, while Modified Demons showed sensitivity to large expansions. Second, the use of only NMI, NCC, and DSC provides a limited view of registration quality; additional metrics such as Hausdorff distance, surface‑based Dice, or Jacobian determinant analysis would capture local geometric fidelity and deformation regularity more comprehensively.

A notable limitation is the reliance on synthetic data generated by ImSimQA. While this approach allows precise control over deformation parameters, it does not replicate the noise characteristics, imaging artifacts, or heterogeneous tissue properties encountered in real CT or MR datasets. Consequently, the reported performance may be optimistic relative to clinical scenarios. Moreover, the study excluded newer deep‑learning‑based DIR techniques, which have shown promising accuracy and speed in recent literature. Future work should extend the evaluation to patient images, incorporate non‑linear and anisotropic deformations, and compare traditional algorithms with state‑of‑the‑art neural network models.

In conclusion, the authors demonstrated that all four DIR algorithms can accurately recover images after modest point displacements (4 mm) and moderate volume scaling (±12 %). However, algorithmic performance diverges when the deformation involves larger expansions, with Modified Demons being the least reliable in preserving volume geometry. These findings underscore the necessity of pre‑clinical validation of DIR tools under application‑specific deformation conditions, and they provide a practical benchmark for selecting appropriate registration algorithms in adaptive radiotherapy workflows.