acceptodds
Under review as a conference paper at ICLR 2027

GED: Gaussian Expected Deformation for Mask-Free Damaged-Section Registration

Abstract

Folds and tears in serial-section electron microscopy create abrupt motion changes between neighboring tissue regions. We present Gaussian Expected Deformation (GED), a pairwise registration method that learns the spatial influence of local displacement fields without damage masks. Anisotropic Gaussian supports can extend along damage boundaries and narrow across them, reducing the mixing of incompatible motions. Sequential RANSAC extracts multiple motion groups from feature correspondences to initialize the fields. Per-pair optimization then refines their parameters and supports through image-similarity and correspondence objectives. On the synthetic benchmark, GED reduces valid endpoint error from 2.09 to 0.215 pixels relative to GPO, a reduction of 89.7%. Near-damage error falls from 6.66 to 1.10 pixels, a reduction of 83.4%. Controlled support ablations and learned-weight visualizations link anisotropy to lower near-damage error and cross-boundary mixing.

Primary area:
applications to neuroscience & cognitive science
Keywords:
serial-section electron microscopy, deformable image registration, damaged-section registration, Gaussian primitives, motion decomposition, feature matching, dense deformation field, connectomics

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_kHhwOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

GED addresses pairwise registration of serial-section electron microscopy images containing folds and tears without requiring damage masks. ALIKED/LightGlue correspondences are decomposed into motion-consistent groups using sequential RANSAC. Accepted matches activate grid anchors, each generating several local displacement primitives. Anisotropic Gaussian supports, decoded partly through shared MLPs, determine how the local fields contribute to a normalized weighted correction field. The parameters are optimized separately for each image pair using photometric, SSIM, and correspondence losses.

The paper derives an idealized two-Gaussian transition-width relation and evaluates registration using dense synthetic endpoint error, instance overlap, image similarity, and manually annotated real-image landmarks. Of 700 synthetic pairs, 40 are used for design and 660 for the primary evaluation. GED reports mean valid EPE of 0.215 pixels versus 2.09 for GPO, and near-damage EPE of 1.10 versus 6.66. Five real pairs provide additional evidence. Support and initialization ablations examine anisotropy and multiple-motion recovery. This is per-pair optimization of an adaptive deformation representation, rather than a pretrained deformation predictor or Neural ODE model.

Critical strengths
  1. The core support mechanism is both well motivated and supported by a controlled ablation. Anisotropic Gaussian supports can extend along a damage boundary while narrowing across it, addressing the mixing of incompatible tissue motions. Using the same raw correspondence cache, the isotropic variant increases near-damage EPE from 0.987 to 1.252 pixels, approximately 27%, and leakage from 0.019 to 0.023 (Table 5). Figure 3 connects the learned support shapes, opposite-side influence, and displacement profile. This is direct evidence that the proposed support geometry contributes beyond the choice of feature matcher. Achieving this without an input damage mask is a meaningful practical advantage.

  2. Substantial and reasonably broad geometric evidence supports accepting the method as a useful contribution. On 660 primary synthetic pairs, GED reduces valid EPE from 2.09 to 0.215 pixels and near-damage EPE from 6.66 to 1.10 relative to GPO. The paper reports results across two source datasets, both damage types, and multiple severities, and keeps 40 design pairs separate from the primary comparison. Instance overlap and manually annotated landmarks on five real pairs complement the dense synthetic errors. GED's ability to handle the difficult DOLW7 fold is particularly valuable, although its gains are not uniform across real cases. Together, these results support practical effectiveness rather than an improvement in image similarity alone.

Critical weaknesses
  1. The headline baseline gap cannot be attributed entirely to anisotropic support learning. Weighted local transformations are established in Expected Affine, while GPO and GaussianDIR already use Gaussian deformation primitives. Appendix B uses different matching and filtering pipelines for EA, GPO, and GED, and Table 5 shows a large benefit from multiple-motion initialization. The anisotropy ablation provides credible evidence for the new support design, but only on the design set. Matched-correspondence comparisons and a primary-set support ablation would strengthen attribution. I regard this as a limitation on the explanation of the full improvement, rather than a reason to dismiss the observed performance of the complete method.

  2. Validation supports a narrower scope than general damaged-section registration. Synthetic damage consists of piecewise translations across approximately straight boundaries, with smooth TPS controls evaluated separately. The five real pairs and 16 landmarks per pair provide useful supporting evidence but limited coverage of combined nonrigid motion, complex boundaries, or annotation variability. The mean real-data gap to GPO is dominated by DOLW7, and GED is not best on SS2L30 or ZW7L18. Claims should therefore remain focused on the tested pairwise setting. The missing explicit local-field expression and normalization conditions also need clarification. These concerns warrant a cautious recommendation, but the existing mechanism evidence and geometric gains are sufficient for weak acceptance.

Questions for authors
  1. Exact local field and initialization: Please give the complete expression for d_i(x), including the action of phi_i and the coordinate origin of its linear term. If an anchor cell contains matches from different motion groups, how are the initial displacements and residual parameters assigned? This is important for reproducing the claimed separation of opposing motions.

  2. Evaluation independence: Are the design and primary evaluation pairs disjoint at the underlying source-section level? Were all real-image manual landmarks excluded from fitting and parameter selection? Please clarify whether Appendix E's “same 80 correspondences” refers exclusively to evaluation landmarks, and report available annotation-repeatability information.

  3. Attribution of the improvement: Do matched-input results exist in which GPO or EA uses the same accepted correspondences as GED, where compatible? Can the anisotropic/isotropic comparison be reported beyond the 40 design pairs with paired uncertainty? The existing ablation supports anisotropy; these checks would clarify how much of the total baseline gap is due to support geometry versus motion initialization.

  4. Normalization coverage: Under what conditions does Eq. (3)'s transition-width argument apply to Eq. (2)'s epsilon-clamped, tile-dependent aggregation? Please clarify whether low total support or candidate changes cause attenuation or tile-boundary artifacts on valid tissue, and give any available coverage diagnostics.

1 comment

Thanks for the replies and extra experiments, they clear up most of my concerns and make the results more convincing. Although the limited real-data evaluation is still worth keeping in mind, I feel more comfortable with the work and trend to keep my initial rating.

Reviewer_BkQuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose Gaussian Expected Deformation for damaged serial-section electron microscopy, a mask-free deformation representation that combines anisotropic Gaussian supports with correspondence-guided Gaussian anchors and sequential RANSAC. They generate a synthetic dataset from CREMI and AC4 sections and evaluate their method on this dataset and five real pairs from the DamagedSection benchmark. On the synthetic benchmark, GED reduces valid endpoint error from 2.09 to 0.215 pixels and near-damage error from 6.66 to 1.10 pixels, while achieving the lowest mean landmark error on the real section pairs (1.472).

Critical strengths

One critical strength is its impressive results, with reductions of 89.7% in valid endpoint error and 83.4% in near-damage error on the synthetic benchmark, while achieving the lowest mean landmark error on the five real section pairs. These results suggest the effectiveness of GED in improving registration accuracy.

Critical weaknesses

One critical weakness is that this paper relies largely on a synthetic dataset for validation. The folds and tears are generated using piecewise translations and may not fully represent the complex physical damage found in real images. Meanwhile, for the real data, the authors use manually annotated landmarks. Although landmark selection criteria are described, further details on how landmarks are selected objectively could improve the reproducibility of the evaluation. Another weakness is GED's relatively high computational cost (68.5 seconds per pair), given that it is a pairwise optimization method.

Questions for authors

How were the landmarks selected and annotated to ensure consistent and reproducible evaluation, independently of the registration results? Also, could the piecewise translation-based generation of synthetic folds and tears favour GED's deformation assumptions?

Reviewer_qx5kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors propose a method for non-diffeomorphic registration in which a clean reference image is registered to a warped image that includes a rip or fold. The method works by first obtaining a set of corresponding points in both images and use them to learn a smooth map based on unisotrophic gaussians centered at the regions containing the correspondences. Results show that the method is able to learn the deformation on synthetic data and evaluation on real data shows that the method performs well on real-world datasets, and outperforms existing methods.

Critical strengths

The real-section results show the results obtained on synthetic data transfer to real data: on DOLW7, GED reaches 2.3 px landmark error where GPO and GaussianDIR are at ~88 px and EA at 4.1 px. Five pairs is thin evidence, but the failure mode of the Gaussian baselines on this case is exactly the one the paper targets. Thus, the results indicate that the method performs well on a difficult real-world task.

Critical weaknesses
  1. Missing reasoning. The authors focus on the description of the method rather than a scientific reasoning for it. This is systematic: the contribution list reads like a checklist of components and does not state what the core novelty is. The core reasoning behind the new algorithm is not described. The related work section focusses on mentioning related work but not on how the current work and its core insights actually relate to it. In particular, it never analyzes why GPO, which also propagates displacements from Gaussian control nodes, should fail on damage, although it is the headline comparison.

  2. Disjoint method description. The method is split between the main text and Appendix B in a way that leaves the main text incomplete and disjointed. The local displacement field d_i(x), is never written down; its “linear residual controlled by a scalar” cannot represent the similarity transforms it is said to be initialized from, since the vector-field is in 2D. The “Cluster + RANSAC” variant in Table 5 is never defined.

  3. Evaluation does not isolate the contribution. The method has a lot of hyperparameters, that are poorly explained. Esepcially, the loss weights and the four(!) learning-rate groups spanning 5e-3 to 2e-1 are stated without any indication of how, or on which data, they were chosen, so the reported comparison cannot be assessed. The baselines are not given the same inputs: per B.2, GPO receives raw matches while GED receives sequential-RANSAC inliers. The authors’ own ablation (Table 5) shows that matching is the dominant factor (near-damage EPE 21.7 vs 0.99), while the titled contribution, anisotropic supports, accounts for about 27% (1.25 vs 0.99). Without GPO run on the same inlier set, the gains in Table 1 cannot be attributed to the proposed representation.

Questions for authors

Can you provide results of GPO using the same sequential RANSAC features? This would strengthen this paper considerably.

3 comments

I think the authors for their reply.

I see that point 1 is partially resolved and point 3 is answered. I gave the paper another read and I maintain that point 2 is not resolved. While the method description now appears to be complete, the writing and description of the method has changed little. Section 3.3 can only be understood after reading 3.4 and reading the appendix is still crucial to understand the main method. Paragraphs are still disjointed, there is no clear red thread. This is likely because the text appears to be in large parts AI generated. This is fine, as long as there is enough human supervision that the full text has a flow and a clear logical thread, something that current gen models still lack and it is still a genuinely human task to clearly lay out the key ideas and how everything connects.

I maintain my score.

Thank you for reading the paper again and for your guidance on the quality of our manuscript.

We agree with your assessment. In the previous revision, we added the missing definitions but kept the text as a sequence of component descriptions. In this revision, we rewrote the Introduction, Related Work and Section 3.

In the Introduction, we now derive the method from a property of the deformation in damaged sections before naming GED. We state the core novelty in the contributions and relate prior work to it in Related Work.

In Section 3, every key term is now introduced before it is used, so Section 3.3 no longer depends on Section 3.4. The parts of the method that had been left in Appendix B are now in Section 3.

The changes are marked in GED_iclr2027_diff.pdf in the supplementary material.

In the opinion of this reviewer, a proper human-guided reformulation of the presented work takes significantly longer than 16hrs and would constitute a major revision. This reviewer does not see the value in continuing being this guide, especially as during this process only singular examples of pervasive problems are resolved without actually attacking the fundamental issues of the writing.

All questions have been answered, my main point stands.

Reviewer_Ajn9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Gaussian Expected Deformation (GED), a method for registering damaged serial-section electron microscopy images without requiring damage masks. The method uses sequential RANSAC to identify different motion groups from feature correspondences. It then represents the deformation using local displacement fields weighted by learnable anisotropic Gaussian supports, aiming to reduce motion mixing across folds and tears. The deformation parameters are optimized for each image pair using image similarity and correspondence losses. Experiments on 660 synthetic image pairs and five real image pairs show improved registration accuracy compared with existing methods. The authors also conduct ablation studies to evaluate the effects of different components.

Critical strengths

First of all, the paper addresses an important problem in serial-section electron microscopy registration. Using anisotropic Gaussian supports to handle different motions across damage boundaries is a reasonable idea. The transition-width analysis also helps explain how the method works.

Another strength is the significant improvement in registration accuracy. As shown in Table 1, valid EPE decreases from 2.09 to 0.215 pixels compared with GPO, while near-damage EPE decreases from 6.66 to 1.10 pixels. The method also achieves lower landmark errors than GPO on all five real image pairs. The ablation study further supports the benefit of using anisotropic Gaussian supports.

Critical weaknesses

First, I think the novelty of the proposed method needs further clarification. GaussianDIR already uses anisotropic Gaussian supports, and the baselines do not appear to use the same motion initialization as GED. Table 5 also shows that sequential RANSAC contributes significantly to the improvement. A fairer comparison with matched initialization would be helpful.

Another concern is that the synthetic experiments mainly use piecewise translations across nearly straight damage boundaries. More complex nonrigid deformations combined with damage are not evaluated. Although the five real cases are useful, 16 landmarks per pair may not be sufficient to evaluate complex deformation patterns.

Finally, the description of the local displacement field is not clear enough. Section 3.4 does not provide its exact mathematical expression. It is also unclear how different motion groups are assigned to the four Gaussian primitives when a grid cell contains correspondences from both sides of a damage boundary.

Questions for authors

First, how would GaussianDIR and GPO perform if they used the same correspondences and motion initialization as GED? This would help clarify the contribution of the proposed deformation model.

Second, have the authors tested more complex deformations, such as curved damage boundaries or nonrigid motion within damaged regions? Would the proposed method still perform well in these cases?

Finally, could the authors provide the exact equation for the local displacement field? How are different motion groups assigned to Gaussian primitives when they appear in the same grid cell?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.