acceptodds
Under review as a conference paper at ICLR 2027

BridgeUQ: Bayesian Aleatoric Uncertainty Quantification For Time-Series Image Registration via Brownian Bridge Prior

Abstract

This paper presents \em BridgeUQ, a novel Bayesian framework for aleatoric uncertainty quantification in pretrained deterministic registration models. In contrast to previous approaches that mainly focus on pairwise registration, our work is the first to model posterior uncertainty over spatiotemporal deformation trajectories that captures the intrinsic ambiguity of registration solutions across both space and time. To achieve this, we first introduce a novel hierarchical Gaussian process prior in the temporal deformation space, based on a Brownian bridge formulation with an inverse-Gamma hyper prior on the diffusion variance, to govern stochastic sampling around the predicted registration solution. We then define an energy-based likelihood function derived from the pretrained registration objective to ensure the sampled deformations remain faithful to the observed data. To facilitate efficient inference, we further develop an amortized maximum-a-posterior (MAP) scheme combined with conditional diffusion sampling to approximate this resulting posterior. It is important to note that BridgeUQ is a plug-and-play module that can be seamlessly integrated with arbitrary pretrained registration models without retraining. We validate BridgeUQ on widely used registration networks using comprehensive cardiac and brain magnetic resonance imaging datasets. Experimental results show that our approach is able to produce anatomically meaningful uncertainty estimates for a variety of learning-based registration models. The proposed BridgeUQ provides a practical and principled pathway toward uncertainty-aware and safety-critical medical image registration.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Uncertainty, Time-series registration, Byesian, Brownian Bridge

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_9B8tOpenReview ↗
Rating
4(+2)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BridgeUQ, a post-hoc framework for estimating aleatoric uncertainty in pretrained time-series image registration models. Stochastic perturbations around predicted velocity trajectories are assigned a Brownian-bridge prior with spatially varying diffusion variance and an inverse-Gamma hyperprior. An energy-based likelihood incorporates the registration objective, while inference alternates variance MAP estimation and velocity sampling.

Critical strengths
  1. A relevant focus on uncertainty across complete deformation trajectories. The paper addresses an important distinction between estimating uncertainty independently for image pairs and representing dependencies across a time sequence. Centering perturbations on an existing prediction also provides a useful route for investigating uncertainty without modifying the original registration network’s parameters.

  2. An explicit and interpretable temporal prior. The Brownian-bridge formulation makes the assumed temporal covariance and endpoint behavior transparent, while the spatial variance field allows different anatomical locations to have different perturbation magnitudes. This is more interpretable than introducing unstructured noise without specifying its temporal organization, although the suitability of the particular assumptions requires further validation.

  3. Encouraging uncertainty-ranking results across multiple backbones and data settings. BridgeUQ achieves lower temporally averaged nAUSC than the reported alternatives on both real-data settings for all four tested backbones. These experiments provide evidence of practical utility beyond a single network configuration.

Critical weaknesses

Before discussing the critical weaknesses, I would like to emphasize that I find the paper sufficiently novel and of high enough overall quality to warrant acceptance. However, the experimental results do not fully substantiate all of the claimed advantages. If my concerns are adequately addressed, I would be willing to increase my score.

1. The variance update is not clearly consistent with the stated conditional posterior.

The method alternates between MAP estimation of the diffusion variance σ\sigma, conditioned on the current velocity trajectory vv, and posterior sampling of vv. However, Eq. (11) includes ∇σE(v)\nabla_{\sigma} E(v), which should vanish when vv is held fixed under the stated model. It is unclear whether this is simply an unsimplified zero term or whether the implementation differentiates through a dependence v(σ)v(\sigma), which would require a different explanation of the optimization objective. This ambiguity does not by itself establish an implementation error, but it needs to be resolved to assess the validity of the inference procedure. Moreover, using a point estimate of σ\sigma rather than integrating over its posterior introduces an approximation whose impact on uncertainty estimates is not established in the main manuscript.

2. The evaluation does not sufficiently establish calibrated, temporally joint uncertainty.

The principal metrics (per-frame nAUSC, Dice, and Jacobian statistics) provide useful evidence about uncertainty ranking and registration quality, but do not directly assess the correctness of cross-time dependencies. Likewise, the correlation between per-frame variance and registration loss in Figure 4(a) does not establish accurate temporal covariance. For example, independently permuting sample identities across time preserves each frame's sample distribution while disrupting the joint trajectories. Ranking metrics are also unchanged by positive rescaling of uncertainty scores, so they cannot establish calibration of uncertainty magnitude. Consequently, additional controls and calibration measurements are needed to substantiate the paper's central claim of a meaningful spatiotemporal posterior, rather than only informative marginal uncertainty maps.

Questions for authors
  1. What objective is actually optimized during the diffusion-variance update? When updating p(σ∣v,I,v^)p(\sigma \mid v, I, \hat{v}), is the current velocity trajectory vv held fixed? Under the stated factorization, the registration-energy contribution ∇σE(v)\nabla_{\sigma} E(v) in Eq. (11) should then be zero. Is this term zero in the implementation, or do gradients propagate through a reparameterized trajectory v(σ)v(\sigma)?

  2. How does replacing the diffusion-variance posterior with a MAP estimate affect uncertainty quantification? Please clarify which posterior distribution the alternating procedure is intended to approximate. Have you assessed the effect of fixing σ\sigma, rather than marginalizing or sampling it, on the resulting uncertainty estimates?

  3. Can you directly evaluate probabilistic calibration and joint trajectory uncertainty? On the synthetic sequences with known ground-truth deformations, could you report empirical coverage at several nominal credible levels, together with interval widths?

6 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer's valuable comments. Please see our point-by-point responses below.

R3.1 What objective is actually optimized during the diffusion-variance update? When updating p(σ∣v,I,v^)p(\sigma|v,\mathcal{I},\hat{v}), is the current velocity trajectory vv held fixed? Under the stated factorization, the registration-energy contribution ∇σE(v)\nabla_\sigma E(v) in Eq. (11) should then be zero. Is this term zero in the implementation, or do gradients propagate through a reparameterized trajectory v(σ)v(\sigma)?

We thank the reviewer for this observation. Here we confirm that the current velocity trajectory vv is held fixed when optimizing over σ\sigma, making the gradient of the first term with respect to σ\sigma zero. No gradients are propagated through the reparameterized velocity trajectory. To avoid this confusion, we have removed this term from Eq. (11) in the revised manuscript.

R3.2 How does replacing the diffusion-variance posterior with a MAP estimate affect uncertainty quantification? Please clarify which posterior distribution the alternating procedure is intended to approximate. Have you assessed the effect of fixing σ\sigma, rather than marginalizing or sampling it, on the resulting uncertainty estimates?

The alternating procedure is intended to approximate the posterior in p(v∣σ,I;v^)p(\mathbf{v}\mid \boldsymbol{\sigma},\mathcal{I}; \mathbf{\hat{v}}) defined in Eq. 12.

Please see the Fig. 4(b) of the main paper regarding the effect of fixing σ\sigma on the uncertainty map. We have also included the quantitative results for the fixed-variance in Table 1. As shown in the table below, applying a fixed variance introduces the same level of perturbation at every voxel regardless of the underlying anatomical structure, which results in uncertainty maps that are less anatomically meaningful. As a result, the AUSC values are close to 1, indicating that the estimated uncertainties have limited discriminative power.

Table 1. Effect of fixed diffusion variance for BridgeUQ$_{\mathrm{TLRN}}$ on the cardiac MRI dataset.

Variancet=2t=2t=4t=4t=8t=8t=10t=10t=12t=12
0.51.021.330.951.550.93
11.120.990.951.391.26
21.011.101.431.211.43

R3.3 Can you directly evaluate probabilistic calibration and joint trajectory uncertainty? On the synthetic sequences with known ground-truth deformations, could you report empirical coverage at several nominal credible levels, together with interval widths?

We thank the reviewer for this insightful suggestion. We agree that evaluating probabilistic calibration on synthetic data with known ground-truth deformations provides a more direct assessment of the uncertainty estimates. We have therefore added an evaluation of trajectory-level uncertainty calibration at multiple nominal credible levels (50%, 80%, 90%, and 95%). Specifically, we construct simultaneous credible bands [1] for the deformation trajectories using the maximum standardized deviation across all temporal frames and both displacement components. We then calculate empirical simultaneous coverage, defined as the proportion of evaluated voxels for which the entire ground-truth deformation trajectory falls within the corresponding credible bands. We also report the average band width of the uncertainty estimates.

Table. 2 show that our Brownian bridge prior achieves empirical coverage closest to the nominal levels across all evaluated credible levels, while producing substantially narrower credible intervals than the two baselines. This demonstrates that our method provides better calibrated uncertainty estimates, achieving a more favorable balance between calibration and informativeness.

Table 2. Empirical trajectory coverage at various nominal credible levels, reported as coverage/interval width.

Temporal prior50%80%90%95%
Brownian motion0.53/1.060.73/1.320.81/1.480.86/1.62
Gaussian prior0.49/0.660.69/0.820.76/0.920.80/1.00
Brownian bridge (ours)0.51/0.770.77/0.950.84/1.050.91/1.13

References

[1] Sigrunn H Sørbye and Håvard Rue. Simultaneous credible bands for latent gaussian models. Scandinavian Journal of Statistics, 38(4):712–725, 2011

AuthorsOpenReview ↗

Reminder: Author-Reviewer Discussion Period

Dear Reviewer,

As the author-reviewer discussion period ends on November 18th, we kindly remind you that we will be unable to communicate after that date. We would appreciate your feedback on whether our response has addressed your concerns. If you have any further questions or need additional clarification, please let us know before the discussion period concludes.

Thank you for your attention.

Best regards,

The Authors

Thank you for the clarifications and additional experiments. I am raising my score to 3 (Weak Accept).

However, Q2 concerns uncertainty in the estimated variance field itself, not whether variance differs across voxels. A spatially varying variance field can still be fixed at its MAP estimate. Your constant-variance ablation demonstrates the benefit of spatial adaptation, but does not assess the effect of ignoring the variance field’s posterior uncertainty.

To address Q2, please compare the MAP-based approach with sampling or marginalizing the variance posterior, for example on synthetic data, and assess changes in trajectory coverage and interval widths. Alternatively, a convincing justification that the MAP approximation has negligible impact would suffice. If this concern is resolved, I will raise my score further.

Reviewer_51NkOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents a Bayesian framework for uncertainty estimation in medical image registration tasks, specifically for time-series. It models the uncertainty over spatiotemporal trajectories using a Brownian bridge formulation of the stochastic noise around the optimal trajectory. Langevin sampling is used to sample stochastic trajectories and coupled to MAP estimates of the diffusion variance, extending previous work. A novelty of the paper is to propose to amortize MAP, resp. Langevin sampling, via a network prediction, resp. a conditional diffusion model. The method is applied to cardiac time-series and longitudinal brain registration, both on synthetic deformations and real data. It evaluates agreement between voxel error and voxel uncertainty, as well as registration field quality in terms of accuracy (Dice) and smoothness (Jacobian det), showing rather strong results compared to relevant baselines.

A specificity of the approach is that it builds on third-party frameworks to provide the initial pointwise estimate of deformations, and focuses exclusively on uncertainty quantification. Given the significant challenges in uncertainty quantification for 3D velocity fields, further amplified in a time-series context, this is in fact a rather welcome and practical choice.

Critical strengths

In my opinion, the paper is suitable for a venue such as ICLR.

There is significant novelty, on a highly technical problem (time-series 2D+t/3D+t registration of medical images), approached through a suitable AI-based framework (Brownian bridges). The paper makes specific choices that hopefully (cf. weaknesses section) make the approach reasonably practical, such as working from a third-party pointwise estimate of the displacement (from one of several high-quality registration frameworks), amortizing the inference, etc. The focus on time-series adds novelty compared to a more traditional 3D setting, and has many practical applications in medical imaging (such as partly demonstrated in the paper).

The paper is overall clear and well structured, despite a relatively high technicity.

The evaluation is sound, on representative cardiac and brain applications, including metrics that assess:

  • the correlation between error and uncertainty (nAUSC)
  • registration accuracy through auxiliary segmentation metrics (Dice)
  • deformation smoothness (Jacobian determinant) all in line with the literature and relevant to the task. Results are qualitatively and quantitatively promising at the very least.

Ultimately, the paper is interesting and describes an approach that I will be keen to try if I investigate time-series registration.

Critical weaknesses

Reproducibility: at this point in time, I was not able to access the anonymous github through the link provided in the paper. Not making the code public would make the work fairly difficult and time-consuming to reproduce given its technical nature, so in this specific case I would count it as a weakness.

I will most likely consider updating my rating from weak acceptance to strong acceptance if the (anonymous) github link becomes accessible, which would strengthen my confidence that it will indeed be made public upon acceptance.

No other critical weakness.

Questions for authors

Could authors please clarify whether the prior over the stochastic noise ηti\eta_t^i in some way encourages spatial smoothness of these stochastic perturbations? My understanding is that there is no such mechanism. How is the apparent regularity of the resulting stochastic velocity (judging from the uncertainty maps) induced?

Could you also please clarify if your implementation of the spatiotemporal perturbations ηti\eta_t^i prevent from integrating the velocity with the scaling and squaring strategy, or whether it is perfectly compatible?

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer's valuable comments. Please see our point-by-point responses below.

R 2.0 Regarding the anonymous github link.

We thank the reviewer for bringing this issue to our attention. We have verified that the anonymous code repository is now accessible with this link https://anonymous.4open.science/r/Bridge_UQ-55BD/. We apologize for the earlier access difficulty and have updated the link in the revised manuscript.

R 2.1 Regarding the clarification whether the prior over the stochastic noise ηti\eta_t^i in some way encourages spatial smoothness of these stochastic perturbations?

Although the prior on the stochastic noise does not explicitly impose spatial smoothness, the sampled velocity fields inherit spatial regularity through two mechanisms. First, because the perturbed velocity, which includes an explicit spatial regularization term (Eq. 3) to penalize spatially irregular velocity fields. Second, the preconditioning matrix used during Langevin sampling was designed to incorporate a Gaussian spatial smoothing kernel. This introduces spatial correlation into the sampling updates and further promotes smoothness in the sampled velocity fields. Additional details are provided in Supplementary Sec. 2. To summarize, spatial regularity is not imposed by the noise prior alone but arises from the combined effects of the spatial regularization in the likelihood and the spatially correlated Langevin sampling updates.

R 2.1 Regarding the clarification whether the implementation of the spatiotemporal perturbations ηti\eta_t^i prevent from integrating the velocity with the scaling and squaring strategy, or whether it is perfectly compatible?

The proposed perturbations are fully compatible with scaling-and-squaring integration. Specifically, BridgeUQ perturbs the velocity field directly, vti=v^ti+ηtiv_t^i=\hat{v}_t^i+\eta_t^i, such that each sampled vtiv_t^i remains a valid velocity field with the same representation as the backbone prediction v^ti\hat{v}_t^i. Each perturbed velocity field is then integrated into a deformation using the same scaling-and-squaring procedure as the pretrained backbone, without requiring any modifications to the integration scheme.

AuthorsOpenReview ↗

Reminder: Author-Reviewer Discussion Period

Dear Reviewer,

As the author-reviewer discussion period ends on November 18th, we kindly remind you that we will be unable to communicate after that date. We would appreciate your feedback on whether our response has addressed your concerns. If you have any further questions or need additional clarification, please let us know before the discussion period concludes.

Thank you for your attention.

Best regards,

The Authors

Many thanks to the authors for addressing my comments and questions. In light of this and of the insights obtained from your responses to other reviewers, I will update my rating.

AuthorsOpenReview ↗

Thanks for increasing the Score!

Dear Reviewer 51Nk,

We are pleased to hear that our response addresses your comments and questions. We sincerely appreciate your thoughtful reconsideration in updating the score from 3 (weak accept) to 4 (clear accept). Thank you for taking the time to review our paper—incorporating your suggestions has undoubtedly improved the clarity of our work.

Best regards,

The Authors

Reviewer_SQ92OpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

BridgeUQ attaches a posterior over velocity trajectories to a frozen registration network. A pretrained model supplies a velocity sequence, uncertainty is a perturbation drawn from a Gaussian process whose covariance is a Brownian bridge. The bridge variance σ(x) is a positive spatial field with an inverse-Gamma hyperprior and does not depend on time. An energy-based likelihood, exp(−E(v)/γ), reweights trajectories by the usual registration objective. Inference alternates a maximum-a-posteriori update of σ with Langevin sampling. For test-time speed, a U-Net is trained to regress σ, and a conditional diffusion model is trained to imitate the Langevin samples. At test time the uncertainty map is the voxel-wise standard deviation of the diffusion samples.

The method is attached to VoxelMorph, TransMorph, TLRN, and LTMA and compared with DIF-VM, PULPo, Rbr, SGMCMC, and Wang et al. Data are a 64×64 synthetic sequence, cardiac cine MRI, and ADNI. The reported criterion is normalized area under the sparsification curve, plus Dice and the fraction of nonpositive Jacobian determinants of the sampled deformations.

Critical strengths

The modeling target is well chosen. Pairwise probabilistic registration produces an independent uncertainty map at each frame. A bridge prior is a direct way to impose a trajectory that starts and ends at a reference, which matches a cardiac cycle.

The finite-dimensional bridge statement in Proposition 1 is standard and the Markov kernel in Equation 6 appears to have the correct shrinkage and matching conditional variance. Centering this kernel on a frozen prediction, and treating the registration energy as a likelihood, is a coherent plug-in around a deterministic network.

The empirical section is broader than a single backbone demo. Amortization moves cardiac inference from 6.4 minutes to 53 seconds per subject with almost unchanged Dice, folding rate, and nAUSC. Across both cardiac cine and ADNI, and for every backbone, the paper reports a lower average nAUSC than the five baselines, with folding held near zero. The temporal-prior ablation is informative: the bridge variance tracks the cardiac registration-loss curve (Spearman 0.85), while Brownian motion and an independent Gaussian prior correlate negatively (about -0.45 and -0.47), as expected when error returns to a small value at the end of the cycle.

Critical weaknesses

The inference equations do not match the model. Under the Markov kernel and the energy likelihood, EE depends on vv and not on σ\sigma, so the ∇σE\nabla_\sigma E term in the printed σ\sigma-gradient is zero. That gradient also does not type-check: βI−α~σ\beta I - \tilde{\alpha}\sigma mixes an identity matrix with a spatial field, and the printed denominator is σ⊤σ\sigma^\top\sigma, a single scalar. The inverse-Gamma score has to be applied element-wise. With T−1T-1 Gaussian log-determinant terms folded into the prior, the element-wise score consistent with the written model is

∂∂σlog⁡p(σ∣v)=∑t(vt−μt∣t−1)22 at σ2+β−α~ σσ2,at=T−tT−t+1,α~=α+T+12.\frac{\partial}{\partial \sigma}\log p(\sigma \mid v) = \sum_t \frac{(v_t-\mu_{t|t-1})^2}{2\,a_t\,\sigma^2} + \frac{\beta - \tilde{\alpha}\,\sigma}{\sigma^2}, \qquad a_t = \frac{T-t}{T-t+1}, \qquad \tilde{\alpha} = \alpha + \frac{T+1}{2}.

The printed velocity score is a different model again. It is the discrete Laplacian

∇vtlog⁡p=−1γ∇vtE+σ−1(ηt+1−2ηt+ηt−1),ηt=vt−v^t.\nabla_{v_t}\log p = -\frac{1}{\gamma}\nabla_{v_t} E + \sigma^{-1}\bigl(\eta_{t+1}-2\eta_t+\eta_{t-1}\bigr), \qquad \eta_t = v_t-\hat{v}_t.

That is the score of a random walk. The bridge kernel contributes the shrinkage (T−t)/(T−t+1)(T-t)/(T-t+1), and vtv_t also enters the next conditional mean. Neither term appears. Langevin dynamics with this score do not target the posterior defined by the proposition. The diffusion model is trained on those samples, so amortization preserves the discrepancy. The optimization is also initialized at σ0∼N(0,I)\sigma_0\sim\mathcal{N}(0,I), outside the inverse-Gamma support (0,∞)(0,\infty), and a set of SS velocity samples is assigned to the next vv with the sample index written as NN. How that set is reduced to the single trajectory required by the σ\sigma update is unspecified.

The temporal uncertainty pattern is largely chosen by the prior. Because η0=ηT=0\eta_0=\eta_T=0 and σ\sigma is constant in time, prior variance along a trajectory is the bridge parabola

Var(ηt)=σ t(T−t)T,\mathrm{Var}(\eta_t) = \sigma\,\frac{t(T-t)}{T},

which is zero at the endpoints and maximal at mid-interval. On cine data that parabola lines up with end-systole, which is what the Spearman ablation measures. For ADNI the paper creates the missing return trip by reversing the longitudinal sequence. The clinical endpoint is then pinned to zero perturbation, and the reverse half is not an observed trajectory. The dataset section says four ADNI time points and 40 cardiac frames. The tables use T=7T=7 and T=15T=15. In the ADNI block, frame tt and frame T−tT-t have identical means and, for most methods, identical standard deviations (DIF-VM at t=1t=1 and t=5t=5 are both 0.335±0.0710.335\pm 0.071). That is the reversed sequence entered twice. Averaging nAUSC over those frames double-counts the early visits.

The baseline table does not compare uncertainty estimators on a common registration. DIF-VM, PULPo, Rbr, SGMCMC, and Wang et al. each estimate their own deformation. BridgeUQ reports Dice and nAUSC for samples around TLRN or LTMA, so a lower nAUSC can come from the backbone. The cardiac gap from the best baseline average (0.592) to BridgeUQ-TLRN (0.531) is smaller than the reported standard deviations, and no paired test is given. On ADNI, Wang et al. at 0.332±0.2280.332\pm 0.228 and BridgeUQ-TLRN at 0.308±0.0490.308\pm 0.049 are not separated. The abstract's claim that BridgeUQ is integrated without retraining covers the frozen registrar only. The usable system trains a U-Net and a conditional diffusion model on Langevin targets. Non-amortized inference remains several minutes per subject. The claim to be the first spatiotemporal deformation posterior is broader than the related work that is discussed.

The real-data accuracy and uncertainty numbers are the wrong measurement. On cardiac cine and ADNI, registration quality is reported as the Dice similarity of a label propagated by the sampled warp, and nAUSC ranks voxels by disagreement between that propagated label and a manual segmentation. Both are tissue-overlap scores. Rohlfing (IEEE Trans. Med. Imaging, 31(2):153–163, 2012) compared intensity similarity and region overlap with landmark target-registration error and found the surrogates unreliable: overlap of large structures can stay high while anatomical points are wrong by several millimeters, and a gain in Dice or in image similarity need not mean a gain in correspondence. A count of nonpositive Jacobian determinants measures regularity, not accuracy. The synthetic sparsification curve is a true velocity error, but it is computed on 64×6464\times 64 smoothed noise of variance 0.01.

This undercuts two claims. The Dice column cannot show that sampled trajectories preserve registration accuracy. The real-data uncertainty ranking has the same flaw. Overlap error concentrates on boundaries of large regions. Misregistration inside a homogeneous structure often changes no labels, so it contributes nothing to Dice or to the segmentation residual. The displayed uncertainty maps place the highest uncertainty in those homogeneous regions and lower uncertainty on boundaries. The metric used to validate the map is blind exactly where the map claims to be informative.

For the brain experiments, point error should replace Dice. The AFIDS release (Taha et al., Scientific Data, 2023, doi:10.1038/s41597-023-02330-9) provides 32 anatomical fiducials on 132 subjects from four structural MRI cohorts and on 14 templates, with human placement error of about 1–2 mm. Correspondence is the Euclidean distance between a propagated fiducial and its curated coordinate,

TRE=∥ϕ(xAFID)−xAFID⋆∥2,\mathrm{TRE} = \bigl\| \phi(x_{\mathrm{AFID}}) - x^{\star}_{\mathrm{AFID}} \bigr\|_2,

which does not depend on the registration objective. The same pretrained backbones, with and without BridgeUQ, should be compared by mean and maximum fiducial error on these pairs. Uncertainty should be read at the fiducial coordinates and checked against that error, by correlation or by sparsification. A method that wins Dice and loses fiducial error has not been shown to register the brain more accurately, and an uncertainty map that is large where fiducial error is small is not calibrated.

AFIDS covers structural brain MRI, including subject-to-template pairs. It does not provide cardiac trajectories or longitudinal ADNI landmarks, so it does not repair the cine half of the results table or the reversed-timeline ADNI analysis. Those still need an observation of correspondence that is independent of the warp. Proofs, step sizes, (α,β,γ)(\alpha,\beta,\gamma), and the train–test split are not in the manuscript. The sparsification sentence says that voxels are removed by increasing uncertainty, that is from least to most certain. Those are opposite orders. The plotted curve is the usual one, with error falling as voxels are removed.

Questions for authors

See above

8 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer's valuable comments. Please see our point-by-point responses below.

R1.1 The ∇σE\nabla_{\sigma}E should vanish.

Please see our response to R3.1.

R1.2 Concerns regarding σ\sigma’s dimensionality, bridge contributions in the posterior score, and sample indexing.

In our formulation, σi(x)\sigma^i(\mathbf{x}) is not a scalar but a d×dd\times d diagonal matrix at each voxel (See Page 4 of the main paper).

The sampled trajectory comes from the posterior distribution, p(v∣σ,I;v^)p(\mathbf{v}\mid\boldsymbol{\sigma},\mathcal{I};\hat{\mathbf{v}}), rather than the Brownian-bridge prior defined in Proposition 1. The second term in Eq. (12) gives the final formulation after reorganizing the bridge kernel and the terms involving vtv_t. The detailed derivation is provided in the revised Supplementary Section 2.

As clarified in the paper, we apply a softplus activation to σ\sigma in both Alg.~1.1 and the amortized network to enforce its positiveness. We apologize that the index NN in Alg.1.1 (line 7) is a typo for SS. The σ\sigma-update uses the gradient in Eq.11 averaged over all SS Langevin samples. We corrected the typo and state this explicitly in Alg.1.1.

R1.3 Regarding temporal timepoints, ADNI’s mirrored sequence and double-counts visits.

We clarify that the dataset section describes the original data, whereas the tables report sequence lengths after preprocessing. For cardiac cine MRI, we uniformly resample the original 40-frame cycle to T=15T=15 frames spanning the full cycle.

For ADNI, the return half is a constructed extension, not additional observed clinical visits. The final observed clinical visit, t3t_3, lies at the midpoint and is not constrained to have zero deformation.

We report the nAUSC for the entire trajectory. Averaging nAUSC over the first three follow-up visits yields the same method ranking.

R1.4 The baseline does not compare uncertainty on a common registration estimator. Lower nAUSC can come from the backbone.

We would like to clarify that a comparison using identical fixed registration predictions is not directly feasible with existing methods, as they jointly estimate deformations and their associated uncertainty. The closest comparison in our experiments is between DIF-VM and BridgeUQ({\text{VM}}), as both are built upon a VM backbone. As shown in Tab. 2, BridgeUQ$_{\text{VM}}$ outperforms DIF-VM on nAUSC on cardiac and ADNI brain dataset.

The improvement in uncertainty estimation is not limited to the backbones. As shown in Tab.2, BridgeUQ$_{\text{VM}}$ also achieves lower average nAUSC than DIF-VM on both the cardiac MRI and ADNI brain datasets, despite using a similar VM architecture.

R1.5 No statistical significant test is given.

We perform a paired tt-test between BridgeUQ$_{\text{TLRN}}$ and the strongest baseline, Wang et al., using subject-level, frame-averaged nAUSC, pairing the two methods’ results for each subject. The improvement is statistically significant on both datasets (p=1.1×10−5<0.05p=1.1\times10^{-5} < 0.05 for cardiac MRI and p=2×10−8<0.05p=2\times10^{-8} < 0.05 for ADNI).

R1.6 Regarding low noise variance in synthetic experiments, limitations of real-data metrics and evaluation for AFIDS landmarks.

According to the suggestion, we re-ran synthetic experiments (Fig. 2) use known ground-truth deformations generated from smoothed Gaussian noise at a higher level, N∼(0,I)N \sim (0, I), to enable voxel-wise error evaluation. Additionally, reliable ground-truth landmark annotations are generally unavailable in real-world longitudinal imaging datasets. Previous works often use Dice similarity and percentage of negative Jacobian determinants as an indirect measure of registration quality. However, we did not intend to claim that improvements in Dice establish improved anatomical correspondence in the absence of ground-truth landmarks. Rather, these experiments assess registration quality, deformation regularity, and the ability of uncertainty estimates to identify potentially unreliable regions using available anatomical annotations.

We further note that validating voxel-wise registration uncertainty on real clinical data remains challenging, as ground-truth deformation fields are generally unavailable. We therefore use complementary evaluations: synthetic experiments with known ground-truth deformations for direct voxel-wise error assessment (Fig. 2), and anatomical segmentation-based metrics for indirect evaluation on real data. We have clarified these limitations in the revised manuscript.

Regarding AFIDS dataset, it provides fiducial annotations for pairwise structural MRI registration. Thus, evaluating BridgeUQ on AFIDS would not provide a direct validation of our primary time-series results.

R1.7 Regarding the sparsification sentence and implementation details.

We thank the reviewer for this, ``increasing'' is a typo, and voxels are removed from least to most certain. Please see Supplementary for these details.

Thanks to the authors for their response. My evaluation remains unchanged. The response fixes the parts of the variance update. The energy term is removed, the velocity is held fixed while the variance is updated, the variance is a positive diagonal field, and the sample index is corrected, I don't see a corrected velocity score. The printed gradient is still a random-walk Laplacian, without the bridge shrinkage (T-t)/(T-t+1) or the dependence of the next conditional mean on the current velocity. Without this, it is unclear that samples come from Proposition 1.

The experimental reason also stands. The ADNI return half is confirmed as a constructed extension, and a paired test against Wang et al. is claimed without the p-values. Uncertainty is still not compared on one fixed registration field except for the VoxelMorph pair. Fiducial error measurements are not included, these are important because Dice and segmentation overlap as used can stay high while anatomical points are wrong.

AuthorsOpenReview ↗

Follow-up discussion

We thank the reviewer for reading through our responses.

  • First of all, please note that the conditional transition coefficients and the dependence of the next conditional mean on the current velocity {\bf are explicitly derived in the Supplementary Material (Eqs. (5)–(13))}. As shown in the steps, the bridge-specific coefficients cancel when the contributions from the two adjacent transitions are combined, which offers the final expression reported in our main paper. As a result, the resulting prior-gradient term is derived from the Brownian bridge transition model, rather than from a random-walk prior. Both the bridge shrinkage and the dependence of the next conditional mean on the current velocity are fully accounted for.

  • Second, regarding the comparison of uncertainty estimates on a fixed registration field, we have provided results using the same frozen VM registration network, thereby ensuring a controlled comparison of uncertainty estimation methods under a fixed registration field. Please note that the existing registration methods do not naturally provide uncertainty estimates. Extending such an evaluation to additional architectures would therefore require adapting their original formulations to incorporate uncertainty quantification as used in our baselines. Given the time constraints of this revision, a systematic evaluation across additional architectures is beyond the scope of the current work and will be considered in future studies. The paired tt-test between BridgeUQ$_{\text{TLRN}}$ and the strongest baseline, Wang et al., and the improvement is statistically significant on both datasets (p=1.1×10−5<0.05p=1.1\times10^{-5} < 0.05 for cardiac cine MRI and p=2×10−8<0.05p=2\times10^{-8} < 0.05 for ADNI).

  • Third, regarding fiducial error measurements, we would like to clarify that these evaluations were performed on the synthetic data, for which ground-truth anatomical correspondences are available. Therefore, we are hoping the concern that Dice and segmentation overlap may remain high despite inaccurate anatomical point correspondences was well addressed through the synthetic-data evaluation. For the real-data assessment, this is a practical limitation when the ground-truth deformation is not available, and we have included this as limitation discussion in the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.