Inverse‐consistent deformable image registration to propagate biomechanical constraints for online adaptive radiotherapy Authors: T J W Draper, C Zachiu, B W Raaymakers Published: 2026 ABSTRACT AbstractBackgroundAccurate motion estimation remains a key challenge in adaptive radiotherapy (ART). Voxel‐wise accuracy is required, as registration errors can compromise the treatment quality. Locally constrained image registration is a promising technique to further improve ART treatment, but relies heavily on quick and accurate delineations of the local areas used to constrain. In online clinical scenarios, these delineations are often available on pre‐treatment image volumes, but not on the online data due to accuracy and time constraints.PurposeWe developed a deformable image registration (DIR) framework that integrates inverse consistency with local biomechanical constraints, enabling tissue‐specific motion modeling using contours from only a single image volume. This approach enforces tissue‐specific constraints on the inverse deformation field and propagates them to the forward deformation through inverse consistency optimization, yielding anatomically plausible motion estimates for ART workflows.MethodsThe proposed framework was evaluated in three scenarios relevant to ART: abdominal CT‐CT and multi‐modality CT‐MR registration with incompressibility constraints on the liver and kidneys, as well as thoracic CT‐CT registration with local rigidity constraints on the ribs and vertebrae. We also investigated the robustness of the framework to image degradation by evaluating motion estimates under added Gaussian noise and artificially induced streaking artifacts. Performance was compared to a baseline DIR model and an unconstrained inverse‐consistent model. Accuracy and anatomical consistency were assessed using landmark alignment, contour overlap, and physics‐based motion analysis.ResultsWe demonstrate that local biomechanical constraints imposed on the inverse deformation field are successfully propagated to the forward motion field through inverse consistency optimization. The proposed framework improves the biomechanical plausibility of the resulting deformation fields at a small cost of landmark‐based accuracy, and maintaining comparable results in terms of contour‐based evaluation. The method is applicable to both mono‐ and multimodality registration and is more robust to noise and artifacts that appear from image acquisition and reconstruction.ConclusionsThe proposed framework uses inverse consistency to propagate local constraints between motion fields. This work demonstrates the use of local biomechanical constraints for DIR in several use‐cases for ART. This methodology requires delineations of local tissue on only one of the image volumes. This approach can become beneficial for applications where voxel‐wise accuracy and anatomical consistency are required, but quick and accurate delineations of new anatomies are unavailable, such as online ART.Keywords: biomechanical constraints, deformable image registration, image‐guided radiotherapy, inverse consistency INTRODUCTION Deformable image registration (DIR) is an essential component in image‐guided adaptive radiotherapy (ART), enabling accurate alignment of images acquired at different times or using different imaging modalities. In ART workflows, the patient anatomy may change due to tumor response, organ movement, or tissue deformation. These anatomical changes correspond to different mechanics underlying tissue‐specific properties, such as compressibility and elasticity. 1 , 2 It necessitates registration methods that not only align image intensities but also explicitly model and constrain deformation behavior to ensure anatomically plausible estimates. Recent guidelines and recommendations for DIR in ART therefore emphasize the anatomical plausibility of the estimated motion, 1 , 2 , 3 , 4 , 5 , 6 particularly in applications such as contour alignment and propagation of quantitative information such as dosimetric values, CT numbers, or functional imaging data (e.g., diffusion and perfusion parameters). These applications require DIR methods with a high degree of accuracy and precision, which simultaneously need to be robust to noise and imaging artifacts. In areas with poor image contrast, constraining the deformation field in line with the underlying physical properties introduces additional prior information into the registration model, which has shown to improve registration accuracy in those areas. 1 , 2 , 3 , 4 Incompressibility is frequently used in DIR to model tissue that can change in shape, but do not change in volume. In these methods, the Jacobian determinant plays a central role, as it quantifies the relative volume change per voxel as estimated from the motion field. In practice, we expect incompressible tissues to have a Jacobian closely distributed around one. Typical values of the liver and kidneys have been determined to be in the range 0.8–1.2. 7 Fu et al. 8 proposed a local variant of the iLogDemons algorithm, 9 by registration of incompressible regions (segmented in advance) with one model, while other regions were handled with an unconstrained model. Rohlfing et al. 10 and Dauguet et al. 11 introduced local incompressibility using the properties of B‐spline basis functions for an efficient numerical implementation. Bosma et al. 12 combined contour‐guided registration with a Helmholtz‐Hodge decomposition to enforce incompressibility and study the impact on dose accumulation. Zachiu et al. 13 combined local incompressibility with smoothness constraints in a spatially adaptive variational registration model. A limitation of the previously mentioned locally constrained registration algorithms is the need for contours to specify the constrained regions. In online‐ART, the delineation and manual editing step is typically the most time‐consuming and error‐prone step in the workflow. 14 , 15 In these workflows, it is therefore typically not feasible to obtain the structures required for constrained DIR. A solution would be to rely exclusively on contours defined on the planning image, which can be delineated offline without stringent time constraints. This requires a technical change in the interpolation methods used to apply the motion field, to use a push‐forward approach instead of a pull‐back approach. However, a downside of this method is that it can cause “holes” in the propagated data, which can potentially cause important complications for mapping quantitative information such as CT numbers or dosimetric values. 16 , 17 , 18 , 19 The pull‐back approach does not have these problems, although local constraints require contours defined on the daily image. Alternatively, one could estimate the motion in the opposite direction, calculating the constrained inverse transformation, before inverting the motion field to obtain the deformation in the forward direction. Inverting the estimated deformation vector field after DIR is employed has been used in the context of deformable dose accumulation, though there the observed difficulty is that it is not always possible to derive a well‐defined inverse motion field. 19 The concept of inverse consistency, that is, estimating the forward and inverse deformation field simultaneously, has been used to obtain a well‐defined inverse map and reduce dose mapping errors. 20 , 21 , 22 However, these methods do not guarantee that the estimated deformation field adheres to any anatomically plausible motion patterns. Instead, we propose a model that jointly optimizes both the forward and the constrained inverse transformations by integrating inverse consistency, while simultaneously imposing local biomechanical constraints, into a novel DIR model. By constraining the motion field in one direction, these constraints will be iteratively propagated through inverse consistency. 23 , 24 We demonstrate improved anatomical fidelity of the motion for CT‐to‐CT/synthetic CBCT/MR registration for the liver, kidneys, and rigid structures in the thorax (ribs and vertebrae). Finally, high computational efficiency is achieved through an optimized GPU implementation in CUDA, enabling seamless integration into online adaptive workflows. THEORETICAL METHODS Baseline DIR model DIR is typically formulated as a minimization problem of a cost function, which jointly optimizes an image similarity metric and enforces regularity of the motion field. The estimated deformation field, u, selected from an admissible function space X, should align the moving image I to the target image J, defined on the image domain Ω⊆R3. The baseline DIR model employed here is as follows: This formulation balances optimization of the image similarity with the regularization of the deformation field, weighted by the parameter α>0. This model employs curvature regularization, which benefits from the fact that rigid transformations are not penalized. 25 , 26 An efficient numerical solution is given by the implicit time‐marching algorithm, leveraging the diagonalization of the Laplace operator in the spectral domain 26 , 27 , 28 , 29 : where τ is the step size of the time‐marching algorithm, the force fn=δuSim(I∘(1+un),J) is the variational gradient of the image similarity metric, and the (inverse) discrete cosine transform (DCT) is used to impose homogeneous Neumann boundary conditions. The matrix K(q) is the representation of the curvature regularization in the DCT basis, with frequency vector q: The tuple (N1,N2,N3) denotes the dimensions of the discretized image volume. The baseline model is used to estimate the forward and inverse motion fields independently of each other, see Figure 1 (top row). A schematic overview of an iteration in the proposed method which updates the forward and inverse asymmetrically, minimizing ICE and constraining the inverse motion field using the partition of the liver and kidneys (bottom). The constrained model is validated to an inverse‐consistent, but unconstrained model (ICDIR) and a baseline model which estimates the forward and inverse motion fields independently (top). Enforcing inverse consistency in DIR There are several ways to enforce inverse consistency on the motion fields. A common approach is to directly penalize the inverse consistency error (ICE) in the cost function. 23 , 24 Here, we adopt the method of interleaved optimization as proposed by Guetter et al., 30 which alternates between the forward and inverse deformation field updates. This approach naturally integrates with traditional registration models. The inverse displacement field, v is estimated through its relationship with u: The inverse displacement field can be efficiently computed using a fixed‐point iterative scheme. 31 A relaxed version was used in interleaved optimization and was used here: Using this method, we define an inverse‐consistent DIR model (ICDIR), following the interleaved optimization scheme, which alternatively updates the forward and inverse motion field using the fixed‐point method to switch from the forward to the inverse motion field (and vice versa). A schematic overview of this model is given in Figure 1 (middle row). Introducing biomechanical constraints to the DIR optimization We show that at least two types of biomechanical constraints are compatible with our framework of constraint propagation through inverse consistency: local rigidity and local incompressibility. These constraints are enforced only on the inverse motion field during optimization, which will propagate to the forward motion field by minimization of the ICE. Constrained image registration with local rigidity constraints With local rigidity, the motion estimate is constrained such that the deformation field is described by some rotation matrix and translation vector in part of the domain. A full description of local rigidity constraints compatible with the update described in Equation (2) is given by Draper et al. 29 The numerical solution uses the alternating direction method of multipliers (ADMM), which introduces two auxiliary variables, to estimate the deformation field followed by an update of the rigid transformations belonging to a list of structures passed to the algorithm. If such a list is available, we replace the update step of the inverse motion field v with the ADMM step. This method comes with an additional model parameter, set to κ=108. The local Leray projection for region‐specific incompressibility constraints This section further extends the baseline registration model by constraining soft tissues based on their incompressibility. Incompressibility can be enforced using the Leray projection, which models the deformation as an incompressible fluid. The projection is defined as follows: The projection P(u) satisfies the property that divP(u)=0 throughout the image domain. It has been employed in previous models to constrain the fluid flow to enforce incompressibility. 9 We apply the projection directly on the deformation field, since a divergence‐free motion field is the leading‐order approximation for incompressibility, that is We extend the Leray projection to enforce incompressibility locally. Let C={Ω′,Ω1,…,ΩN} define a partition of Ω, that is, a number of pairwise‐disjoint regions that cover Ω, and let χΩ:Ω↦{0,1} denote the indicator function: Then, local incompressibility is achieved by replacing the target function divu in the potential‐defining Poisson Equation (6), by the piecewise g(u): The local Leray projection satisfies the property Thus, in the region Ω′, this projection does not alter the divergence of the motion field. Figure 2 shows the local Leray projection applied on a synthetic deformation field. The difference maps show that the motion field is altered in such a way that the divergence is unchanged in the background Ω′, although the difference amplitude does show that differences permeate outside the constrained regions to accommodate local incompressibility. The local Leray projection was integrated within our framework by following each update of the inverse motion field by a projection on a list of given incompressible structures, as highlighted in Figure 1 (bottom row), and the numerical algorithms are given in Appendix A. This figure illustrates the effect of the local Leray projection z on a synthetically generated deformation field u. The partition C consists of the white shape (Ω1), the gray circle (Ω2) and the black background (Ω′). The projection enforces zero divergence in Ω1. The difference maps show that the divergence is modified within Ω1 and remains unchanged outside, while changes in the deformation field itself extend beyond Ω1. Inclusion of a second region Ω2 demonstrates that the local Leray projection can also enforce non‐zero divergence values (here set to the mean), as further detailed in the Discussion section. Implementation details All registration algorithms were implemented in C++/CUDA 32 and executed on GPU hardware. As the cuFFT library does not support a native DCT implementation required for Equation (2), a custom implementation was made using the Fast Fourier Transform and appropriate coefficient manipulation based on the work of Makhoul 33 and the multidimensional DCT implementation in Matlab (MIRT_DCTN). Using Matlab's MEX API, the registration models were compiled into Matlab executable MEX functions. Data pre‐ and post‐processing was done in Matlab (version 2025b). All experiments were conducted on a workstation equipped with an Intel(R) Xeon(R) W‐2133 CPU and an Nvidia Quadro RTX 5000 GPU. EXPERIMENTAL METHODS The proposed model was evaluated in four experiments relevant to online‐ART workflows, using three publicly available datasets. Together, these experiments demonstrate the versatility for mono‐modality and cross‐modality applications, as well as multiple types of biomechanical constraints in the abdomen and thorax. An overview of the datasets and experiments is provided in Figures 3 and 4. Abdominal CT‐guided ART with incompressibility constraint on the liver: The first ten cases of the Liver‐DIR‐QA database were used, containing abdominal CT‐CT image pairs with annotated landmarks on vessel bifurcations in the liver. 34 These manually validated landmarks, on average 55 per image volume, were used to validate the accuracy. Manual delineations of the liver were created and used to enforce incompressibility. The local correlation coefficient (LCC) was used as the similarity metric, with Gaussian width σ=5 voxels, and regularization parameters were set to (α,τ)=(5.0,1.0). Abdominal CT‐guided ART with incompressibility constraint on the liver: The first ten cases of the Liver‐DIR‐QA database were used, containing abdominal CT‐CT image pairs with annotated landmarks on vessel bifurcations in the liver. 34 These manually validated landmarks, on average 55 per image volume, were used to validate the accuracy. Manual delineations of the liver were created and used to enforce incompressibility. The local correlation coefficient (LCC) was used as the similarity metric, with Gaussian width σ=5 voxels, and regularization parameters were set to (α,τ)=(5.0,1.0). Thoracic CT‐guided ART with rigidity constraint on the ribs and vertebrae: The POPI database was used to validate the model on six 4DCT image series, binned in ten phases of the respiratory cycle. 35 Of each image set, the maximum inhalation and exhalation phase were selected for registration. Each image volume contains anatomical landmarks annotated inside the lungs, with imposed rigidity constraints on the surrounding ribs and thoracic vertebrae. Segmentations of these structures were obtained using TotalSegmentator. 36 The parameter configuration was the same as for the abdominal CT‐CT registration, besides a smaller Gaussian width of the LCC similarity: σ=2.5 voxels. Thoracic CT‐guided ART with rigidity constraint on the ribs and vertebrae: The POPI database was used to validate the model on six 4DCT image series, binned in ten phases of the respiratory cycle. 35 Of each image set, the maximum inhalation and exhalation phase were selected for registration. Each image volume contains anatomical landmarks annotated inside the lungs, with imposed rigidity constraints on the surrounding ribs and thoracic vertebrae. Segmentations of these structures were obtained using TotalSegmentator. 36 The parameter configuration was the same as for the abdominal CT‐CT registration, besides a smaller Gaussian width of the LCC similarity: σ=2.5 voxels. Noise and artifact robustness for adaptive CBCT‐guided radiotherapy: Robustness of the proposed constraint propagation to noise and imaging artifacts is checked by adding Gaussian noise or streaking artifacts, which are characteristic for cone‐beam CT image reconstruction. The TIGRE toolbox was used to artificially create cone‐beam projection data, which was subsequently undersampled before reconstruction with the Feldkamp‐Davis‐Kress algorithm. 37 , 38 This setup was used to study the effect of cone‐beam artifacts on the quality of the estimated deformations, in terms of landmark alignment and biomechanical fidelity. Both scenarios are illustrated by Figure 4. Parameter configurations are unchanged to study the effect of image degradation. Noise and artifact robustness for adaptive CBCT‐guided radiotherapy: Robustness of the proposed constraint propagation to noise and imaging artifacts is checked by adding Gaussian noise or streaking artifacts, which are characteristic for cone‐beam CT image reconstruction. The TIGRE toolbox was used to artificially create cone‐beam projection data, which was subsequently undersampled before reconstruction with the Feldkamp‐Davis‐Kress algorithm. 37 , 38 This setup was used to study the effect of cone‐beam artifacts on the quality of the estimated deformations, in terms of landmark alignment and biomechanical fidelity. Both scenarios are illustrated by Figure 4. Parameter configurations are unchanged to study the effect of image degradation. Multi‐modality CT‐MR image registration with local incompressibility on the liver and kidneys: We used the eight cases of the abdominal MR‐CT training database from The Cancer Imaging Archive (TCIA), as published during the 2021 Learn2Reg challenge. 39 , 40 , 41 , 42 These image volumes come together with manual segmentations of the liver, spleen, and kidneys, see Figure 3. The liver and kidneys were used to enforce an incompressibility constraint. The second case was removed since both kidneys and liver are only partially visible on the MR‐scan. 28 Also, for case one the left kidney was not evaluated since it was only present in one of the scans. Images were pre‐aligned with an affine transformation, where the scaling ensures that the volumes of the liver and kidneys retain their volume, before DIR was performed. For this multi‐modal application, a patch‐averaged normalized gradient fields (NGF) similarity metric was used with a 5×5×5 patch size and the regularization parameters to (α,τ)=(1.0,1.0). 43 Multi‐modality CT‐MR image registration with local incompressibility on the liver and kidneys: We used the eight cases of the abdominal MR‐CT training database from The Cancer Imaging Archive (TCIA), as published during the 2021 Learn2Reg challenge. 39 , 40 , 41 , 42 These image volumes come together with manual segmentations of the liver, spleen, and kidneys, see Figure 3. The liver and kidneys were used to enforce an incompressibility constraint. The second case was removed since both kidneys and liver are only partially visible on the MR‐scan. 28 Also, for case one the left kidney was not evaluated since it was only present in one of the scans. Images were pre‐aligned with an affine transformation, where the scaling ensures that the volumes of the liver and kidneys retain their volume, before DIR was performed. For this multi‐modal application, a patch‐averaged normalized gradient fields (NGF) similarity metric was used with a 5×5×5 patch size and the regularization parameters to (α,τ)=(1.0,1.0). 43 Overview of image pairs from the Liver‐DIR‐QA (left), POPI (middle), and Abdomen MR‐CT (right) datasets, together with delineations of the liver (blue), ribs (purple), thoracic vertebrae (orange), kidneys (red), and spleen (green), which were used for enforcing local constraints and/or validation of the registration models. Only one set of contours was used for constrained DIR, the other was only used for validation purposes. The figure shows an axial slice of a 3D image volume of the Liver‐DIR‐QA database. Each row, from left to right, shows the deterioration of image quality where either the variance of the Gaussian was increased (top), or the number of synthetically generated projection angles was decreased (bottom). In this figure, σ2 indicates the variance of the (zero‐mean) Gaussian noise, and N the number of projection angles used during reconstruction. All images were resampled to an isotropic voxel size of 2 mm, and normalized between zero and one. After registration, the forward and inverse motion fields were resampled to the original image resolution, and all subsequent analyses were performed on the original resolution. The registration algorithms employed a coarse‐to‐fine resolution strategy. 44 Each registration was analyzed using metrics derived from the motion field itself, using annotated landmarks and using contour‐based metrics when contours are available. Although all methods compute both forward and inverse motion fields (cf. Figure 1), the analysis presented here focuses on the forward motion field, in order to assess how constraints imposed on the inverse field affect the resulting forward deformation. Concretely, incompressibility of the tissue was measured using the Jacobian determinant of the deformation field, with Einc.=∥detJu−1∥ as metric for incompressibility. Rigidity was measured with the residual error of the orthogonality constraint, 45 , 46 Additionally, we calculated the residual inverse consistency: to verify the achieved inverse consistency between the forward and inverse deformation fields. Contours were evaluated in terms of the Dice similarity coefficient (DSC), measuring the overlap before and after registration. Annotated landmarks were propagated with the forward deformation field, and we report on the target registration error (TRE), measuring the distance between corresponding landmark. Throughout the manuscript, we use the ColorBrewer color schemes for visualization of the results, accessible via https://github.com/DrosteEffect/BrewerMap 47 RESULTS Abdominal CT‐CT registration with incompressibility constraints on the liver Figure 5 illustrates the deformable registration results for a representative abdominal CT‐CT image pair. Visually, the baseline and ICDIR models exhibit pronounced local expansions and contractions inside the liver, reflected as hot and cold spots in the Jacobian determinant maps. In contrast, the constrained model produces a much more uniform deformation field, indicating improved regularity. The boxplots below show the distribution of the Jacobian determinant evaluated inside the liver structure, where the constrained model displays a substantially narrower distribution. Qualitative overview of motion estimates from the three registration models (unconstrained, ICDIR, and the proposed method). The first column shows the moving and target image, as well as the overlap (purple = moving, green = target). The following columns show the results from the baseline, ICDIR and constrained model with the Jacobian determinant (top row), registered CT (middle row) and overlap between the target and registered image (bottom row). The Jacobian determinant also shows the outline of the liver that was used to enforce incompressibility in the constrained model. These observations are supported by a quantitative analysis across all cases. When considering the mean Jacobian determinant computed in the liver mask, values ranged between 0.91–1.07 for the baseline, 0.93–1.07 for ICDIR, and 0.99–1.01 for the constrained model. However, since the mean value does not reflect spatial variations within the deformation field, we also include a voxel‐wise analysis of the Jacobian determinant. This is quantified in the second column of Figure 6, where zero variance of Gaussian noise corresponds to the original images. The mean incompressibility error Einc. is reduced by approximately a factor of four compared to the baseline. Results on the Liver‐DIR‐QA dataset. The top row shows the mean TRE, Einc. and EICE as function of the Gaussian noise variance. Each point corresponds to the average value over all ten registrations. All quantities have been calculated inside the liver mask. The bottom row shows the same quantities as function of the number of undersampled projection angles. Figure 6 also shows the effect on the TRE of the anatomical landmarks, and the EOC metric evaluated inside the liver. The mean ICE of both the ICDIR and constrained models are reduced by at least an order of magnitude compared to the baseline model, and are well below the voxel size. The TRE of the landmarks before registration was 14.9(7.3) mm, with the case with the largest initial displacement having a mean TRE of 22.8 mm and the smallest being 4.8 mm. Using the estimated displacement fields, the TRE is reduced to 1.0(1.4) mm (baseline), 1.0(1.1) (ICDIR) and 1.4(1.4) mm (constrained). The baseline model failed to reduce the mean TRE below the maximum voxel dimension (2.5 mm) for one image pair. The computation time was on average 1.7 s for the baseline model (per direction), 3.6 s for ICDIR and 3.9 s for the constrained model. Respiratory motion estimation of thoracic 4DCT with rigidity constraints on ribs on vertebrae Figure 7 shows the results for the thoracic CT‐CT registration. The initial TRE before registration, averaged over the six cases, was 8.0 (5.6) mm, which was reduced to 1.4 (1.3) using the baseline model, 1.9 (2.2) mm using ICDIR and 1.7 (1.9) mm with the constrained model. The TRE was reduced below the maximum voxel size (2 mm) except for one image pair using ICDIR and the constrained model. The mean EOC is reduced by an order of magnitude from the unconstrained model, showing that the constrained model much better satisfies the assumed rigidity in the constrained vertebrae and ribs. Results on the POPI dataset. The top row shows the mean TRE of lung landmarks, the EOC in the ribs and vertebrae and the EICE as function of the Gaussian noise variance. Each point corresponds to the average value over all six registrations. All quantities have been calculated inside the masks of the ribs and thoracic vertebrae. The bottom row shows the same quantities as function of the number of undersampled projection angles. Registration took on average 1.9 s for the baseline model (per direction), 8.2 s for ICDIR and 9.9 s for the constrained model. Robustness to noise and synthetic streaking artifacts in CBCT imaging Figures 6 and 7 further show the relation between the motion estimates and the introduced Gaussian noise or synthetic streaking artifacts. Both the ICDIR and constrained models have mean ICE below 0.2 mm in the liver for all degrees of noise and artifacts in both datasets, well below the voxel size. Secondly, the constrained model shows that the mean Einc. (Liver‐DIR‐QA) and the mean EOC (POPI) are almost completely unaffected by noise, while it does degrade these metrics for the unconstrained models. In terms of the TRE, the results of the models are almost similar when the degree of noise and artifacts is small, but the TRE increases the slowest for the constrained model. Multi‐modality CT‐MR image registration with local incompressibility on the liver and kidneys Figure 8 demonstrates the results of the forward motion estimates obtained from the NGF‐driven registration methods on an image pair from the multi‐modality Abdomen MR‐CT dataset. All registration models resolve the misalignment, most clearly visible in the kidneys and spleen. The baseline and ICDIR models show significant hot and cold spots in the assumed incompressible liver and kidneys. The constrained model demonstrates complete incompressibility inside these structures, while slightly affecting neighboring structures such as the spleen. Multimodal registration from the Abdomen‐MRCT dataset, showing the image pair and overlap before DIR, and after with the baseline mode, ICDIR and the constrained model. The top row shows the Jacobian distribution of the estimated motion, with outlines of the liver, kidneys and spleen, as well as the CT body contour/field‐of‐view. The image overlap shows the CT in purple, and MR in green. Tables 1 and 2 show that these results generalize over the entire dataset. The differences between the motion estimates of the three registration models result in slightly different contour overlap, with mean DSC in the range 0.84–0.86 for the liver, 0.82–0.83 for the kidneys, and 0.75–0.77 for the spleen. The constrained model achieves this alignment with simultaneously significantly lower incompressibility error and standard deviation of the Jacobian distribution. Table 2 shows that ICDIR reduces the Einc. in all structures with respect to the baseline model. Further analysis shows that this happened by reducing the large outliers in the Jacobian to achieve well‐defined invertibility of the motion field. The constrained model further improves incompressibility in the constrained structures (liver and kidneys) by a factor of 3 to 4, while the motion estimate in the spleen remains mostly unaffected. TABLE 1The DSC of the liver, kidneys and spleen, before registration, after affine alignment, and after subsequent DIR with the baseline, ICDIR and constrained model of the Abdomen‐MRCT dataset.Dice similarity coefficientInitialAffineBaselineICDIRConstrainedLiver0.54 (0.15)0.78 (0.07)0.84 (0.08)0.86 (0.07)0.85 (0.07)Kidneys0.33 (0.14)0.58 (0.12)0.83 (0.10)0.82 (0.11)0.82 (0.11)Spleen0.42 (0.20)0.57 (0.14)0.75 (0.13)0.77 (0.14)0.76 (0.15) The DSC of the liver, kidneys and spleen, before registration, after affine alignment, and after subsequent DIR with the baseline, ICDIR and constrained model of the Abdomen‐MRCT dataset. TABLE 2Statistics on the volume variations as estimated by the DIR methods inside the liver, kidneys and spleen of the Abdomen‐MRCT dataset. The left three columns show the mean absolute error of detJu−1, and the latter three columns show the mean and standard deviation (in parentheses) of the Jacobian distribution. Einc. detJu BaselineICDIRConstrainedBaselineICDIRConstrainedLiver0.3080.1850.0371.09 (0.41)1.04 (0.25)1.01 (0.07)Kidneys0.1780.1490.0410.92 (0.21)0.99 (0.20)1.01 (0.07)Spleen0.2740.1650.1691.14 (0.34)1.06 (0.24)1.07 (0.23) Statistics on the volume variations as estimated by the DIR methods inside the liver, kidneys and spleen of the Abdomen‐MRCT dataset. The left three columns show the mean absolute error of detJu−1, and the latter three columns show the mean and standard deviation (in parentheses) of the Jacobian distribution. Registration of the multi‐modality image volumes took on average 6.4 s for the baseline model (per direction), 13.0 s for ICDIR and 13.8 s for the constrained model. DISCUSSION The proposed methods demonstrate clear potential to improve DIR in image‐guided ART, particularly in scenarios where contour availability is limited but physically consistent deformation fields are required. Through simultaneous estimation of the forward and inverse deformations, the approach addresses a key limitation of conventional DIR, which typically do not enforce physically consistent motion constraints. This is particularly relevant in the presence of imaging noise and artifacts, where robustness of the estimation motion is essential, all while maintaining compatibility with the time‐constraints for online‐adaptive workflows. A relevant application is CBCT‐guided online ART, where a major bottleneck remains the manual review and correction of contours propagated from the planning CT to the daily CBCT. 15 The proposed framework directly addresses this limitation by requiring contour delineations only on the planning image, potentially reducing the need for time‐consuming manual adjustments. In parallel, recent advancements in AI‐based segmentation, such as TotalSegmentator and MedSAM, are rapidly increasing the availability of fast and accurate delineations on MR and CT imaging modalities. 36 , 48 A limitation of the proposed approach is that its relative advantage decreases in clinical settings where accurate segmentations are readily available for both the planning and daily images. However, although CBCT image quality has improved through advances in acquisition and reconstruction techniques, the modality remains affected by noise, artifacts and has reduced soft‐tissue contrast, which makes accurate segmentation more challenging than on conventional CT or MR imaging. Consequently, AI‐based segmentation on CBCT cannot yet be considered a solved problem, and robust contour propagation methods remain clinically relevant. Importantly, DIR remains essential for applications that require voxel‐wise correspondence between images, including deformable dose accumulation, propagation of CT intensity values, and propagation of functional imaging parameters. Recent studies have highlighted the importance of complementing traditional registration accuracy metrics with measures of biomechanical plausibility 3 , 4 and inverse consistency 17 , 19 , 20 , 21 , 22 for these applications. The proposed framework therefore provides a method for improving the physical plausibility and consistency of deformation fields while remaining compatible with existing ART workflows. The experimental results provide evidence that inverse consistency can transfer biomechanical constraints between the inverse and forward motion fields. By simultaneously minimizing the ICE and constraining the inverse motion field, we have shown that motion field characteristics such as incompressibility and rigidity are transferable properties. Concretely, for abdominal CT‐CT registration, we observed that the mean Einc. in the liver decreased by a factor four compared to the baseline model, while preserving comparable landmark registration performance. A similar pattern is observed for the thoracic CT‐CT registration task, where the mean EOC was reduced by a factor four compared to the baseline model, resulting in mean EOC comparable to values reported in Draper et al. 29 and Staring et al. 46 For both datasets, the mean TRE was below the maximum voxel size for all methods, which is argued as a sufficient degree of accuracy for radiotherapy applications by the AAPM Task Group 132. 4 However, a reduction in mean TRE was observed in Figures 6 and 7 when constraints were enforced, particularly in the absence of noise and artifacts. Moreover, we observed that these constraints remain consistent when noise and imaging artifacts are introduced to the imaging volumes. From Figures 6 and 7, the evaluated metrics in the abdominal (liver) experiment are more sensitive to induced noise and artifacts than those in the thoracic experiment. We expect this to be the case due to the sharp contrast of ribs, vertebrae, and lung tissue in POPI, compared to the liver bifurcations used for landmark annotation in Liver‐DIR‐QA, which lose contrast with increasing noise and artifacts. This framework for DIR is compatible with different similarity metrics, such as the LCC for mono‐modality, and NGF for multi‐modality image registration as shown in Figure 8 and Tables 1 and 2, achieving contour overlap comparable to the current state‐of‐the‐art registration algorithms. 28 Incompressibility constraints on the liver and kidneys have been enforced to result in a narrow distribution of the Jacobian determinant. The Jacobian determinant of soft tissues with high water content such as the liver and kidneys has been determined to be within the range 0.8–1.2. 7 Using the local Leray projection, our constrained model ensures that almost all voxels are contained within this range, where the baseline and ICDIR model show significant variance in this structures. Our framework achieves this result without affecting the contour overlap of these structures, or surrounding structures (e.g., the spleen). Averaged over these seven cases, the achieved DSC is similar to the current state‐of‐the‐art registration models, 28 although here the affine pre‐alignment aided as a pre‐processing step. Finally, using a GPU‐accelerated implementation, we have shown that inverse‐consistent, constrained DIR can be achieved at the order of seconds, opening up applications within radiotherapy with strict time‐constraints, such as online ART. Figures 6 and 7 and Table 2 show that enforcing inverse consistency (green) already reduces the compressibility of the liver and non‐rigidity of the ribs and vertebrae, respectively, compared to the baseline model (blue). Minimizing inverse consistency ensures that the motion fields have a well‐defined inverse. Since motion fields with a small or negative Jacobian determinant are not invertible, enforcing inverse consistency reduces the extreme values from the Jacobian determinant distribution, leading to a narrower distribution. Additionally, performing constrained registration and inverting the motion field afterwards can therefore lead to ill‐defined behavior as there is no guarantee, even with appropriate regularization, that the motion fields have a positive Jacobian determinant in the image domain. Currently, even state‐of‐the‐art registration algorithms can have a (small) number of negative Jacobian determinants. 49 Alternatively, the motion field could be expanded in an underlying basis of smooth functions, such as (cubic) B‐splines, although those approaches can lead to overly smooth deformation fields. This is highlighted in the context of sliding motion between the rib‐lung interface by Vishnevskiy et al., requiring linear instead of cubic B‐splines for motion estimation. 50 Future research will explore which motion constraints can be propagated using this approach. Clearly, incompressibility cannot be used in all regions, for instance to capture the motion patterns of the lungs, the gastro‐intestinal tract, and the bladder. 51 Tzitzimpasis et al. 52 demonstrated with the generalized div‐curl regularization that assuming a nearly uniform Jacobian determinant, rather than strict incompressibility, improved pulmonary motion estimation. In this spirit, we remark that the local Leray projection can be extended by modifying Equation (9): which sets the divergence in each region Ωi to a user‐defined value μi. This is exemplified in Figure 2, where the local Leray projection simultaneously sets the divergence to zero in one region, while in a second region, it is set to its mean value. This generalization has potential applications when the amount of compression/expansion is known beforehand, or could be used during optimization to smooth out the degree of compression from the motion field. Finally, shearing motion at the rib‐lung interface results in a sharp transition of the motion gradient at the boundary. Development of inverse consistency methods of piecewise smooth motion fields would therefore be beneficial to incorporate within the proposed framework. The class of diffeomorphic registration algorithms inherently produce smooth motion fields with smooth inverses and could therefore be of potential interest. Examples are the ANTs, LDDMM, CLAIRE, and the class of log‐Demons models. 53 , 54 , 55 , 56 However, beyond constraints such as incompressibility, it remains unclear how constraints applied during the estimation process affect the final deformation field, since these methods regularize an intermediate variable (the velocity field) rather than the motion field itself. Future work could therefore focus on incorporating local biomechanical constraints within this class of registration models. CONCLUSION This work proposed a framework for variational DIR that propagates localized biomechanical constraints between the forward and inverse motion fields via inverse consistency. The framework was evaluated on abdominal and thoracic CT datasets, as well as abdominal CT‐MR data, demonstrating that local incompressibility constraints in the liver and kidneys, and rigidity constraints in the ribs and vertebrae can be effectively propagated between deformation fields. This resulted in substantially improved biomechanical plausibility, achieved at a small cost in landmark‐based accuracy and comparable results for contour alignment. Furthermore, this method retained its accuracy in the presence of noise and streaking artifacts representative of CBCT imaging. These findings suggest that the proposed approach provides a practical mechanism to include biomechanical information into ART workflows, when there is a limited availability of delineations. CONFLICT OF INTEREST STATEMENT Our department has a research agreement with Elekta AB on the development of adaptive strategies for the CBCT‐linac. Elekta AB had no role in the preparation, review, or approval of the manuscript and the decision to submit the manuscript. ACKNOWLEDGMENTS This work was supported by the Health∼Holland, Topconsortia for Knowledge and Innovation ‐ Life Sciences and Health (DARE project). NUMERICAL ALGORITHMS 1, 2, 3 The local Leray projection. Fixed‐point iterative method for the inverse of a motion field (Inv). Locally constrained, inverse‐consistent image registration algorithm DATA AVAILABILITY STATEMENT The three datasets used in the analysis of this study are publicly available at https://github.com/open‐vv/popi‐model (POPI), https://zenodo.org/records/13738577 (Liver‐DIR‐QA) and https://learn2reg.grand‐challenge.org/Datasets/ (Abdomen MR‐CT). Further details may be obtained from the corresponding author upon reasonable request. REFERENCES Associated Data Data Availability Statement The three datasets used in the analysis of this study are publicly available at https://github.com/open‐vv/popi‐model (POPI), https://zenodo.org/records/13738577 (Liver‐DIR‐QA) and https://learn2reg.grand‐challenge.org/Datasets/ (Abdomen MR‐CT). Further details may be obtained from the corresponding author upon reasonable request.