acceptodds
Under review as a conference paper at ICLR 2027

ReStack-Mamba: Stage-wise Residual Refinement for Volume Electron Microscopy Registration

Abstract

Volume electron microscopy (vEM) assembles three-dimensional tissue volumes from physically cut sections, but cutting, collection, and imaging introduce rigid motion and nonuniform elastic deformation. The central challenge is to correct these section-specific distortions while preserving long-range axial continuity without propagating errors through the stack. We present ReStack-Mamba, a self-supervised framework for this coupled local-and-global registration problem. A Local Elastic Aligner first estimates in-plane correspondence between neighboring sections. Two independently trained, z-preserving StackMamba stages then use XY, XZ, and YZ context to correct successive residual errors without changing the section order. Training the second stage on errors remaining after the first allows it to learn a distinct correction; robust axial post-processing suppresses isolated outliers. Across six simulated volumes spanning five tissues and one real-world dataset, ReStack-Mamba improves XZ/YZ SSIM by approximately 10% over existing methods and achieves the best aggregate PSNR, 3D NCC, and 3D MI. Per-section and deformation-severity analyses show that the advantage persists through long stacks and becomes more pronounced under stronger distortions. The real-volume result further shows improved axial continuity, supporting robust vEM registration across diverse tissue structures.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
volume electron microscopy, serial-section registration, deformable registration, self-supervised learning, state-space models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ux75OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses 3D registration for serial-section volume electron microscopy (vEM). ReStack-Mamba first applies a Local Elastic Aligner (LEA), a RAFT-style iterative optical-flow estimator, to correct in-plane deformation between adjacent sections. Two independently trained, z-preserving StackMamba stages then serialize the feature volume along the three orthogonal XY, XZ, and YZ planes, model each plane with a separate residual Mamba branch, and adaptively fuse them to predict bounded in-plane residual displacement fields; the second stage is trained on the residual errors of the first stage’s frozen output. Finally, a deterministic axial Gaussian smoothing plus trimmed-mean post-processing suppresses outlying corrections.

Critical strengths
  1. Well-motivated problem decomposition. vEM registration is decomposed into two coupled subproblems—within-section local elastic deformation correction and cross-section axial-consistency refinement—with a clear correspondence between them; Figure 1’s illustration of the error-accumulation mechanism is intuitive and effective.

  2. The z-preserving tri-plane Mamba design fits the task structure. The encoder downsamples only in-plane and retains the full section sequence; the XY/XZ/YZ scans capture within-section morphology and axial discontinuity respectively, and Mamba’s linear complexity suits long section stacks.

  3. Complete and well-controlled experiments and ablations. Six simulated volumes span five tissues, plus one real stack; the evaluation includes a per-section error-accumulation analysis and deformation-severity robustness experiments, with controlled comparisons (identical input stacks, identical contrast windows); the component ablation and the stage-design ablation (including the “Stage 1 applied twice with shared weights” control) are well conceived.

Critical weaknesses
  1. Unresolved objective conflict in the two-stage design: corrections for axial consistency may sacrifice in-plane fidelity, and the safeguards are only soft constraints

The paper decomposes registration into two “coupled” subproblems—in-plane deformation and axial consistency—yet optimizes them with two serial stages, sidestepping the cost of the coupling itself: the second stage targets axial consistency, but its corrections are in-plane residual fields; in ambiguous regions with missing or blurred structures, “axial consistency” can only be achieved at the cost of pairwise fidelity between adjacent sections. The safeguards are exclusively soft constraints—the identity, magnitude, and anchor regularization terms of Eq. (10) and the bounded residual. No loss term explicitly anchors pairwise alignment against being sacrificed, and no experiment quantifies this trade-off. Moreover, these regularizers penalize all corrections indiscriminately, without distinguishing fidelity-preserving from destructive ones, which inherently conflicts with the goal of “learning a correction.” The XY column of Table 2 confirms the risk: after adding Stage 1, in-plane PSNR slightly decreases and in-plane SSIM rises only marginally; Stage 2 leaves it unchanged; the only appreciable in-plane gain comes from the post-processing—which is not a learned module.

  1. The necessity of the second-stage learning module is unproven, and the bulk of the pipeline’s gain comes from non-learned post-processing

The paper’s central narrative is that “the second stage learns residual corrections beyond the first stage,” but the authors’ own controlled experiments do not support it: in the component ablation (Table 2), all metrics remain essentially unchanged after the second stage is added; in the stage ablation (Table 3), the independently trained second stage shows no substantive difference from the naive control of “Stage 1 applied twice with shared weights”; and the bulk of the pipeline’s gain comes from the deterministic smoothing post-processing. “Training two stages” is not evidence that “both stages contribute”—this is attribution by circular argument. Combined with Point 1, the more plausible explanation is that the protective regularizers have sealed off the learnable space: the second stage achieves safety, not effectiveness.

  1. The evaluation protocol is coupled with the smoothing post-processing, casting doubt on the interpretability of the gains

The final step of the pipeline is a smoothing operation along the z axis (Gaussian smoothing plus trimmed mean), while the primary metrics—SSIM/PSNR—are computed on XZ/YZ views, and the undeformed reference volume is naturally smooth along z. Consequently, any z-direction smoothing—regardless of whether it improves alignment—mechanically inflates these metrics. This is not an implementation detail but a design flaw of the evaluation protocol: the metrics reward “smoothness,” and the method happens to smooth; this coupling makes the gains in the main tables uninterpretable as genuine alignment improvement. The authors themselves acknowledge in Appendix S6 that the operation smooths fine texture. Breaking this coupling requires controlled experiments: comparisons with the post-processing removed, with the same post-processing applied uniformly to all methods, or with smoothing-insensitive evaluations (e.g., downstream segmentation benefits). Absent these experiments, the paper cannot answer how much of the gain over vEMRec reflects alignment quality and how much reflects metric bias.

  1. The most direct competing baseline is missing, leaving the method’s position in the SOTA landscape unclear

The paper cites NeuroTrajReg (NeurIPS 2025)—which uses the same OpenOrganelle data and the same SSIM/MI evaluation protocol and is the most direct competing method—yet includes no quantitative comparison; ssEMnet (cited), as well as GaussReg / EMReg, both compared in NeuroTrajReg, are likewise absent. Among the existing baselines, TrakEM2 (2012) and EFSR are clearly weak. Without a comparison against the most direct competitor, readers cannot determine the method’s true position within the current SOTA lineage, and the central claim of “improvement over the strongest baseline” loses its reference point.

  1. The evaluation protocol is set on the method-favorable side, and reporting seems selectively

The overall design—optical-flow LEA first, learned modules handling only small residuals—naturally fits smooth elastic deformation; the evaluation protocol accordingly consists of simulated smooth Gaussian elastic fields, and the deformation-severity experiment merely rescales the same fixed deformation fields, excluding real sectioning artifacts such as folds, tears, missing sections, and intensity drift (the authors acknowledge in Appendix S6 that these are out of scope). The method’s assumptions and the evaluation protocol are isomorphic, forming a circle: validating a method with exactly the deformation type it assumes. Real data comprises a single stack evaluated only qualitatively, so performance under real artifacts cannot be judged. Meanwhile, the main table (Table 1) shows only the three volumes with the largest gains over vEMRec; the other three, relegated to the appendix, show markedly smaller gains, and the abstract’s “approximately 10%” and “best aggregate PSNR” claims rest on this selected subset.

Questions for authors

Please see the weaknesses as shown above.

4 comments
AuthorsOpenReview ↗

Response to Reviewer ux75 (1/3)

W1. Axial consistency and in-plane fidelity

We thank the reviewer for highlighting this trade-off. With final filtering disabled, halving Stage 2 regularization improves XY, axial, and segmentation metrics together. On EPFL (deformation seed 46), initial alignment is fixed and both conditions use revised signed NCC; only the six Stage 2 regularization weights change.

Table W1-1. Fidelity and residual geometry

MetricStage 2 (1×)Stage 2 (0.5×)
XY PSNR (dB)20.4584120.48892
XY SSIM0.469700.47260
Axial SSIM0.48473 ± 0.000030.48786 ± 0.00002
Dice, frozen segmenter0.75182 ± 0.000060.75517 ± 0.00022
Fraction of non-positive residual Jacobian determinants01.571 × 10⁻⁷
Mean absolute residual displacement (pixels)0.076410.14616
XY gradient-magnitude ratio to reference0.327890.32832

These results quantify the gains and geometric changes under weaker regularization.

**Coupled problems and module roles.**Our method assigns the two coupled subproblems to two distinct modules: LEA addresses local in-plane alignment between adjacent slices, while StackMamba addresses axial consistency across the stack. This division corresponds to LEA versus StackMamba, not to Stage 1 versus Stage 2. LEA produces an initially aligned stack; StackMamba then combines XY features with XZ/YZ context to predict bounded XY residuals that correct remaining cross-slice discontinuities. The subproblems remain geometrically coupled because moving one slice in XY changes its correspondence with both neighbors. Both refinement stages belong to StackMamba: Stage 1 refines the LEA output, and independently initialized Stage 2 learns remaining errors on fixed Stage 1 outputs. Overlapping-window predictions are fused into a whole-stack displacement field.

Pairwise similarity and regularization. The original adjacent-slice NCC squared covariance; revised signed NCC distinguishes positive and negative correlation. It directly guides pairwise correspondence, while variance weighting reduces low-texture contributions. Larger residuals and improved metrics after relaxation suggest that the original constraints limited useful correction.

Table W1-2. Common stage-specific weights

Loss termStage 1Stage 2
Adjacent-slice similarity11
Image identity10.5
Displacement magnitude0.50.25
Spatial gradients0.80.4
Boundary/mean-displacement anchoring0.30.15
First-order axial differences0.40.2
Second-order axial differences0.80.4

For stage s∈1,2s\in\\{1,2\\}, IinI_{\rm in} is the input patch, I^\hat I its refined output, and uu the backward residual field. Both stages use:

\mathcal L^{(s)}={}&\mathcal L_{\rm sim}^{(s)} +\lambda_{\rm id}^{(s)}\mathcal L_{\rm id} +\lambda_{\rm mag}^{(s)}\mathcal L_{\rm mag} +\lambda_{\rm grad}^{(s)}\mathcal L_{\rm grad}\\\\ &+\lambda_{\rm anchor}^{(s)}\mathcal L_{\rm anchor} +\lambda_{z1}^{(s)}\mathcal L_{z1} +\lambda_{z2}^{(s)}\mathcal L_{z2}. \end{aligned}$$ $$\mathcal L_{\rm sim}^{(s)}=-\frac12\sum_{\ell\in\\{1,1/2\\}} \frac{\sum_p w_{\ell p}\rho_{\ell p}}{\max(1,\sum_p w_{\ell p})},$$ $$\rho_{\ell p}=\operatorname{clip}\left(\frac{c_{\ell p}}{\sqrt{(v^A_{\ell p}+\epsilon)(v^B_{\ell p}+\epsilon)}},-1,1\right), \quad w_{\ell p}=\operatorname{stopgrad}\left(\frac{v^B_{\ell p}}{v^B_{\ell p}+\epsilon}\right),\quad\epsilon=10^{-4}.$$ $A,B$ are later/earlier adjacent refined slices; $c$ is local covariance; $v^A,v^B$ are variances clamped to non-negative values. $\ell$ indexes scale; $p$ covers valid $9\times9$ windows over all adjacent pairs. Half-scale images use bilinear resizing with aligned corners; stopgrad stops gradients through the variance weight. $$\mathcal L_{\rm id}=\langle(\hat I-I_{\rm in})^2\rangle,\quad \mathcal L_{\rm mag}=\langle u^2\rangle,$$ $$\mathcal L_{\rm grad}=\frac13\sum_{d\in\\{x,y,z\\}}\langle(\Delta_d u)^2\rangle,\quad \mathcal L_{\rm anchor}=\langle u_1^2\rangle+\langle u_D^2\rangle+\langle(\bar u_{zxy})^2\rangle,$$ $$\mathcal L_{z1}=\langle|\Delta_z u|\rangle,\quad \mathcal L_{z2}=\langle|\Delta_z^2 u|\rangle.$$ $\langle\cdot\rangle$ denotes an elementwise mean; $\Delta_d$ is a forward difference. $D$ is patch depth; $u_1,u_D$ are first/last-slice displacements; $\bar u_{zxy}$ is each component's spatial mean. Identity/magnitude constrain image and displacement changes; gradients/axial differences control continuity; anchoring limits boundary and mean drift. All main volumes and deformation strengths share these weights. The revised balance improves in-plane, axial, and downstream metrics together.
AuthorsOpenReview ↗

Response to Reviewer ux75 (2/3)

W2. Independent Stage 2 versus applying Stage 1 twice

We thank the reviewer for distinguishing the use of a second stage from evidence of its contribution. With revised regularization, independent Stage 2 provides a measurable gain over two shared-weight refinement passes on EPFL. The original regularization configuration did not show this advantage.

We fix initial alignment, disable final filtering, and use the same frozen segmenter. Stage 1 is trained for 24 epochs; independent Stage 2 receives 24 additional epochs. The comparison uses EPFL deformation seed 46 and reports means ± sample SD across training initializations 43, 44, and 45. Axial SSIM averages XZ and YZ SSIM.

Table W2-1. Independent versus shared-weight refinement

ConfigurationTraining epochsAxial SSIMDice
Stage 1240.48364 ± 0.000060.75005 ± 0.00009
Stage 1 twice, shared weights240.48491 ± 0.000100.75213 ± 0.00008
Independent Stage 2 (1×)24 + 240.48473 ± 0.000030.75182 ± 0.00006
Independent Stage 2 (0.5×)24 + 240.48786 ± 0.000020.75517 ± 0.00022

Compared with shared weights, revised independent Stage 2 improves Dice by 0.00304 and axial SSIM by 0.00295. Original 1× Stage 2 is slightly below the shared-weight control on both metrics. The increase in residual magnitude from 0.07641 to 0.14616 pixels in W1 further supports the interpretation that the original constraints restricted learnable correction.

Thus, the revised independent stage adds improvement beyond applying Stage 1 twice. This increment is observed across three initializations and remains without final filtering: its contribution is measured by a direct comparison, rather than inferred from the two-stage design alone.

W3. Does final smoothing explain the main results?

We thank the reviewer for pointing out that axial image metrics can reward smoothing. With ReStack-Mamba's final filtering disabled, mean PSNR and SSIM across all six simulated volumes remain higher than vEMRec's. The advantage therefore cannot be attributed entirely to final smoothing.

We disable final Gaussian smoothing, trimmed-mean filtering, and optional displacement-field filtering, while fixing LEA weights and initial-alignment settings. The volumes are jrc_mus_liver, jrc_mus_pancreas, jrc_mus_liver3, jrc_mus_kidney, jrc_mus_heart, and jrc_mus_skin. The table reports equally weighted means across these six volumes, giving one aggregate value per metric for each method.

Table W3-1. Six-volume mean results without final postprocessing

MethodPSNR (dB)SSIMNCCMI
TrakEM221.09160.46110.87250.7912
SEAMLeSS21.17450.46830.87660.8091
EFSR14.55040.33350.39070.2113
vEMRec24.93930.66060.94411.1451
ReStack-Mamba25.17250.70650.95551.1620

Within each volume, PSNR/SSIM are weighted by valid slice counts per view. Archived TrakEM2, SEAMLeSS, and EFSR outputs are evaluated under the common protocol without additional filtering.

Relative to vEMRec, mean PSNR increases by 0.2332 dB, SSIM by 0.0459, NCC by 0.0114, and MI by 0.0169. ReStack-Mamba ranks first on all four aggregate metrics. The EPFL frozen-segmenter evaluation also gives Dice 0.75517 versus 0.74755 for vEMRec (complete comparison in W5), and the stage comparison above retains a Dice increment without final filtering. These downstream results provide evidence beyond image similarity, supporting registration benefits that do not depend on final smoothing.

AuthorsOpenReview ↗

Response to Reviewer ux75 (3/3)

W4. Direct competing baselines

We thank the reviewer for highlighting the limited coverage of the original comparison. We added an adaptation of public NeuroTrajReg code; ReStack-Mamba achieves higher values on all four metrics for the same six inputs. We retrain the adaptation using a reconstructed training pipeline and an 800-pixel field of view, with identical deformed inputs and a common evaluator.

Table W4-1. NeuroTrajReg / ReStack-Mamba on identical inputs

VolumePSNR (dB)SSIMNCCMI
Liver21.2159 / 23.68210.4199 / 0.58660.7788 / 0.87160.7635 / 1.0840
Pancreas21.2104 / 24.46320.4399 / 0.63980.8458 / 0.92840.7943 / 1.1335
Liver322.2694 / 26.44300.5221 / 0.72230.9224 / 0.97180.9898 / 1.3738
Kidney20.4989 / 24.48930.4315 / 0.65620.8728 / 0.95110.7565 / 1.1273
Heart22.4297 / 26.67790.5637 / 0.73510.9548 / 0.98340.8966 / 1.1811
Skin21.1276 / 24.97520.4861 / 0.66450.9288 / 0.97100.7481 / 1.0360

Each pair lists adapted NeuroTrajReg / ReStack-Mamba; bold indicates the better value. Deformation seed: 45. Volume names map to the six jrc_mus volumes listed in W3. PSNR/SSIM are weighted by valid slice counts; NCC/MI are computed over the full volume.

PSNR improves by approximately 2.47-4.25 dB and SSIM by 0.167-0.225 over this adapted implementation.

W5. Real-data validation and volume selection

We thank the reviewer for raising both concerns. We added segmentation on real VNC data without simulated deformation and on EPFL data with mitochondrial annotations under controlled deformation. The full six-volume summary also retains the advantage over vEMRec.

VNC directly tests real sections without adding Gaussian elastic deformation. All six methods use the same frozen segmenter, threshold 0.5, and reference annotations; metrics are computed per slice and then averaged.

Table W5-1. Real VNC segmentation (higher is better)

MethodDiceIoUPrecisionRecall
TrakEM20.553690.405540.554840.67112
SEAMLeSS0.663310.496240.619380.71397
EFSR0.574160.408480.560390.68852
NeuroTrajReg0.668320.501860.613600.70706
vEMRec0.664130.497150.623500.71042
ReStack-Mamba0.689210.511650.630100.72828

Dice/IoU exceed vEMRec by 0.02508/0.01450. Precision and recall also lead all baselines: improved detection accompanies improved prediction accuracy. This extends evidence beyond image similarity under the assumed simulated deformations to downstream performance on a real acquisition sequence.

Regarding volume selection and “approximately 10%” in the abstract, the full six-volume mean relative axial-SSIM gain is 6.62%, with per-volume gains of 2.60%-12.26%. We average XZ/YZ SSIM for each method, calculate each volume's relative improvement over vEMRec, then average equally across volumes. The abstract now reports 6.62%, consistent with all six volumes. Mean PSNR without final filtering is also higher: 25.1725 versus 24.9393 dB (W3), so the aggregate advantage does not rely on selecting three volumes.

Table W5-2. EPFL controlled-deformation segmentation

MethodDiceIoU
TrakEM20.412660.25997
SEAMLeSS0.692230.52932
EFSR0.446560.28746
NeuroTrajReg0.683750.51947
vEMRec0.747550.59687
ReStack-Mamba0.755170.60664

EPFL uses simulated deformation seed 46 on a real image volume, the same frozen segmenter, threshold 0.5, and reference annotations. Dice/IoU improve over vEMRec by 0.00762/0.00977, consistent with VNC.

The additional VNC segmentation evaluation addresses the lack of quantitative evidence on real acquisition data acknowledged in the original Appendix S6. Together with the annotated EPFL experiments and complete six-volume reporting, it strengthens the evidence for practical registration benefits beyond simulated deformations and qualitative observations.

Reviewer_ux75OpenReview ↗

After reading the authors' rebuttal

Thanks for the authors' rebuttal. The following questions are still unclear to me.

  1. W5-1 reports mean IoU of 0.51165 and mean Dice of 0.68921 for ReStack-Mamba. With the same binary masks and slice weights, that mean IoU implies a mean Dice no higher than about 0.67694. Were empty masks or slice weights handled differently, and does the Dice gain over vEMRec remain 0.02508 under a common convention?

  2. On EPFL, the independent Stage 2 outperforms a second pass of Stage 1 at 0.5× regularization but not at 1×. Halving the regularization also increases its mean absolute residual displacement from 0.076 to 0.146 pixels. At comparable mean displacement magnitudes, which errors, if any, does Stage 2 correct that the shared-weight pass misses?

I would maintain my origial socre at this stage.

Reviewer_2XctOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses registration of serial-section vEM stacks, where section preparation and imaging introduce both in-plane elastic deformation and inconsistency along the sectioning axis. The proposed self-supervised pipeline (Sec. 3, pp. 3–5) has four parts. A Local Elastic Aligner (LEA) estimates in-plane displacement between neighboring sections with iterative coarse-to-fine flow and triplet self-supervision, combined with Gaussian-weighted neighbor-flow consensus at inference. Two independently parameterized, z-preserving StackMamba stages then scan features along the XY, XZ and YZ planes and predict bounded (±0.5 px) in-plane residual displacement fields; Stage 2 is trained on the frozen Stage 1 output so that it targets the errors left by Stage 1. Finally, a deterministic inference-time post-processing step (Gaussian smoothing of the Stage 2 residual with σ_z = 0.35 plus two radius-2 trimmed-mean passes) is applied. The method is evaluated on six simulated OpenOrganelle mouse tissue volumes (558–1,387 sections; smooth Gaussian deformation fields, fixed seed; Table S1, p. 12) and one real Kasthuri11 stack, the latter qualitatively only since no undeformed reference is available. Against TrakEM2, SEAMLeSS, EFSR and vEMRec, the method reports the best PSNR, SSIM, 3D NCC and 3D MI on all six simulated datasets (Tables 1, S2–S4), with the advantage over vEMRec growing as the deformation level increases (Fig. 6). Cumulative and stage-design ablations are given in Tables 2–3.

Critical strengths
  1. Decomposing the problem into local elastic correction and stack-level axial refinement is a reasonable abstraction, and the motivation (pairwise error propagation, axial jitter) is well supported by Fig. 1 (pp. 1–3).
  2. The ablations are well designed. The shared-weights double-application control (Table 3, p. 9) is the right way to separate stage-specific training from a second refinement pass, and the per-section curves (Fig. 5) and the deformation-severity sweep (Fig. 6) address the drift and robustness questions I would naturally raise.
  3. The results are consistent: the method is best or near-best on all metrics on all six simulated datasets, and the gap over the strongest baseline (vEMRec) widens with deformation level, reaching 0.0401–0.0432 in mean per-section SSIM at α = 5 (p. 7).
  4. The paper is honest and reproducible. The full training and inference protocol is in the appendix (pp. 12–13), and the Limitations section (S6, p. 19) states explicitly that the real-data evaluation is qualitative only, that the post-processing smooths fine texture, and that vEMRec wins at α = 1. The AI-use statement (p. 10) complies with the policy.
Critical weaknesses
  1. All the quantitative evidence comes from a single-seed simulation protocol (S1.1, p. 12), and the only real dataset is evaluated qualitatively without a reference. No error bars are reported in the paper, so I cannot tell whether small differences, such as the +0.08 dB PSNR on kidney (Table S2, p. 14), are meaningful. The simulated deformation is a smooth Gaussian field, and the paper acknowledges (S6) that real artifacts such as folds, tears and missing sections are outside the model's scope, so the gap between the simulated and the real setting is unquantified.
  2. NeuroTrajReg (Zhang et al. 2025a, NeurIPS 2025) is not included in the comparison. It is the closest recent method: it targets the same task and, according to its published supplementary, it is evaluated on the same six OpenOrganelle datasets with the same Gaussian smoothing scale (σ = 0.08), but at a smaller deformation magnitude (α = 1.0, vs 3–4 here). I would have expected this comparison. Second, the EFSR numbers are substantially lower than those published for EFSR on the same datasets (3D NCC 0.19–0.58 here, Tables 1/S2, vs ≈0.95–0.98 in the NeuroTrajReg paper, α = 1.0, 1184×1184 crops). A difference in deformation severity may explain part of this, but the paper does not discuss the gap, so the EFSR comparison should be interpreted with caution.
  3. There is no report of runtime, memory or throughput, for this method or any baseline. Scalability is central in this application (SEAMLeSS is a petascale pipeline, and vEMRec claims TB-scale throughput), and part of the motivation for using Mamba is efficiency (p. 4). Without these numbers I cannot assess the practical cost of the method relative to vEMRec.
  4. The stage-wise design, which is the main claim of the paper, receives limited support from the ablations. In Table 2 (p. 7), Stage 2 changes XZ SSIM by only 0.002 (0.6095 to 0.6115), and YZ SSIM drops slightly (0.6118 to 0.6112); most of the remaining gain comes from the deterministic post-processing (to 0.6380/0.6386). In Table 3 (p. 9), independent Stage 2 training improves XZ SSIM by 0.0024 over the shared-weights variant while 3D NCC decreases (0.9708 to 0.9683). The post-processing parameters (σ_z, trimming radius, number of passes) have no sensitivity analysis, which matters because post-processing is the second-largest contributor in Table 2. In addition, the abstract's "approximately 10%" SSIM gain over existing methods exceeds the measured gains on most datasets (2.5%–12.6%, p. 14).
Questions for authors
  1. Why is NeuroTrajReg, which is evaluated on the same OpenOrganelle datasets with the same σ = 0.08 smoothing, not included in the comparison? If it is prior work from your group, please clarify the relationship between the two methods.
  2. EFSR achieves 3D NCC of 0.19–0.58 under your protocol, while the published NeuroTrajReg paper reports ≈0.95–0.98 for EFSR on the same datasets (α = 1.0). Where did your EFSR implementation come from, which configuration did you use, and can you break down how much of the difference comes from the deformation level?
  3. Please report runtime, peak memory and per-section throughput on one common dataset, for your method and for vEMRec and SEAMLeSS.
  4. Could you report mean ± std over at least three deformation seeds? In particular, is the +0.08 dB margin on kidney within the variance?
  5. Can you provide a quantitative evaluation on a real dataset with an available reference (for example FemFlyBrain, used in NeuroTrajReg)? I would also like to see a sensitivity analysis of the post-processing parameters.
3 comments
AuthorsOpenReview ↗

Response to Reviewer 2Xct (1/3)

Q1. NeuroTrajReg comparison and technical relationship

We thank the reviewer for pointing out this omission. The original comparison prioritized the more recently published vEMRec; we have now added NeuroTrajReg, and ReStack-Mamba achieves higher values on all four metrics across the same six inputs. The shared OpenOrganelle data and smoothing scale make NeuroTrajReg a directly relevant comparison.

We adapted and retrained its public code using a reconstructed training pipeline and an 800-pixel field of view. Both methods receive identical deformed inputs and are assessed with the same evaluator.

Table Q1-1. NeuroTrajReg / ReStack-Mamba on identical inputs

VolumePSNR (dB)SSIMNCCMI
Liver21.2159 / 23.68210.4199 / 0.58660.7788 / 0.87160.7635 / 1.0840
Pancreas21.2104 / 24.46320.4399 / 0.63980.8458 / 0.92840.7943 / 1.1335
Liver322.2694 / 26.44300.5221 / 0.72230.9224 / 0.97180.9898 / 1.3738
Kidney20.4989 / 24.48930.4315 / 0.65620.8728 / 0.95110.7565 / 1.1273
Heart22.4297 / 26.67790.5637 / 0.73510.9548 / 0.98340.8966 / 1.1811
Skin21.1276 / 24.97520.4861 / 0.66450.9288 / 0.97100.7481 / 1.0360

Each cell lists the public-code adaptation of NeuroTrajReg followed by ReStack-Mamba, using deformation seed 45. Bold indicates the better value in each pair. The volumes correspond to jrc_mus_liver, jrc_mus_pancreas, jrc_mus_liver3, jrc_mus_kidney, jrc_mus_heart, and jrc_mus_skin. PSNR/SSIM are weighted by valid slice counts; NCC/MI are computed over the full volume.

Both methods address serial-section alignment, but their deformation models differ. NeuroTrajReg uses neural ordinary differential equations to model continuous deformation trajectories. ReStack-Mamba predicts bounded in-plane residual fields from sequential features in orthogonal views and composes the displacement fields of two stages. These differences distinguish their deformation representations and optimization paths.

The direct comparison shows that ReStack-Mamba's advantage extends beyond vEMRec to NeuroTrajReg under identical inputs and the stated adaptation protocol. The results correspond specifically to the reconstructed training pipeline and 800-pixel field of view described above.

Q2. EFSR implementation and deformation strength

We thank the reviewer for requesting that implementation and input effects be separated. With the EFSR implementation and weights fixed, increasing deformation strength from 1 to 3 reduces NCC by 0.2039-0.3891 across five volumes. Stronger deformation is therefore a quantifiable contributor to the lower scores in our evaluation.

The archived adaptation uses EFSR's public code, frozen FE_serial weights, and inference with context. Image dimensions are padded to multiples of 64, with the first and last observed slices used as endpoints. The following comparison keeps the implementation and weights fixed on 800 × 800 inputs and changes only deformation strength. The main experiments use α=3\alpha=3.

Table Q2-1. EFSR deformation-strength comparison with fixed implementation and weights

VolumeNCC: α=1NCC: α=3Change
jrc_mus_kidney0.49950.1701-0.3294
jrc_mus_heart0.92410.5921-0.3320
jrc_mus_pancreas0.55510.2605-0.2945
jrc_mus_skin0.63190.4280-0.2039
jrc_mus_liver30.60460.2155-0.3891

For example, heart NCC decreases from 0.9241 to 0.5921, and liver3 NCC from 0.6046 to 0.2155. This comparison isolates deformation strength from implementation changes and shows that the stronger perturbations used in the main experiments substantially reduce this adapted method's performance.

Thus, scores across papers must be interpreted in light of deformation strength. This fixed-implementation experiment provides direct evidence for the contribution of stronger perturbations to lower NCC.

AuthorsOpenReview ↗

Response to Reviewer 2Xct (2/3)

Q3. End-to-end runtime, memory, and throughput

We thank the reviewer for suggesting a common computational-cost comparison. On identical inputs and hardware, ReStack-Mamba has lower mean end-to-end runtime than vEMRec, with slightly higher peak allocated GPU memory; SEAMLeSS has the highest throughput.

We test all three methods on the same A100-PCIE-40GB using the same 100 slices of 800 × 800 pixels and identical timing boundaries. Each method is warmed up once and timed three times, with execution order rotated. Timing includes checkpoint loading and output I/O, but excludes common input preparation and Python startup. Times are means ± sample SD.

Table Q3-1. Full-pipeline computational cost on identical inputs

MethodTotal time (s)Time/slice (s)Throughput (slices/s)Peak allocated GPU memory (GiB)
vEMRec55.75 ± 0.100.5571.7940.272
SEAMLeSS12.56 ± 0.310.1267.9640.293
ReStack-Mamba51.61 ± 3.440.5161.9370.299

Relative to vEMRec, ReStack-Mamba reduces mean runtime by 4.14 s (approximately 7.4%) and increases throughput from 1.794 to 1.937 slices/s, while allocated memory increases from 0.272 to 0.299 GiB. SEAMLeSS takes 12.56 ± 0.31 s and processes 7.964 slices/s.

The accuracy gains over vEMRec therefore do not entail longer end-to-end runtime for this workload: a small increase in allocated memory accompanies a modest reduction in mean runtime. Although SEAMLeSS is faster, ReStack-Mamba achieves better reconstruction quality. We therefore consider the additional computational cost justified by the improvement in registration accuracy.

Q4. Stability across three deformation seeds

We thank the reviewer for asking us to assess the stability of small improvements. We have added paired statistics across three deformation seeds. The revised model produces positive refinement gains on all six volumes, with a kidney PSNR gain of 0.1393 ± 0.0013 dB relative to initial alignment. This gain varies little across the tested deformations.

Using deformation seeds 45, 48, and 49, we fix the model weights and initial-alignment settings and disable final postprocessing. For each seed, we calculate the paired difference between complete refinement and initial alignment. The table reports means ± sample SD across these three deformations.

Table Q4-1. Complete refinement minus initial alignment across three deformation seeds

VolumeΔPSNR (dB)ΔSSIMΔNCCΔMI
Liver0.0820 ± 0.00090.00632 ± 0.000420.00242 ± 0.000020.00253 ± 0.00020
Liver30.1589 ± 0.00130.00598 ± 0.000050.00112 ± 0.000010.00894 ± 0.00011
Kidney0.1393 ± 0.00130.00525 ± 0.000030.00163 ± 0.000030.00663 ± 0.00003
Heart0.1804 ± 0.00190.00582 ± 0.000020.00071 ± 0.000010.00999 ± 0.00005
Pancreas0.1087 ± 0.00340.00622 ± 0.000070.00184 ± 0.000060.00464 ± 0.00019
Skin0.1308 ± 0.00090.00484 ± 0.000050.00098 ± 0.000010.00868 ± 0.00003

The kidney PSNR gain ranges from 0.1379 to 0.1400 dB, substantially exceeding its variation across seeds. All four mean gains are also positive for the other five volumes. These results support consistent refinement benefits from the revised model across the tested deformations, rather than an improvement confined to one deformation seed.

AuthorsOpenReview ↗

Response to Reviewer 2Xct (3/3)

Q5. Real-data quantification and postprocessing effects

We thank the reviewer for requesting quantitative real-data evaluation and analysis of postprocessing. We added segmentation evaluations on real VNC data without additional simulated deformation and on EPFL data with mitochondrial annotations under controlled simulated deformation. ReStack-Mamba leads the six registration methods on all four VNC metrics and on EPFL Dice/IoU. Its image-similarity advantage also remains with final filtering disabled.

Access to the suggested FemFlyBrain data requires an application to the original authors. Our application has not yet been approved, so we used VNC for additional real-data quantification. All six methods use the same frozen segmenter and threshold 0.5. VNC metrics are computed against reference annotations per slice and then averaged.

Table Q5-1. Real VNC segmentation (higher is better)

MethodDiceIoUPrecisionRecall
TrakEM20.553690.405540.554840.67112
SEAMLeSS0.663310.496240.619380.71397
EFSR0.574160.408480.560390.68852
NeuroTrajReg0.668320.501860.613600.70706
vEMRec0.664130.497150.623500.71042
ReStack-Mamba0.689210.511650.630100.72828

The respective gains over vEMRec are 0.02508, 0.01450, 0.00660, and 0.01786, providing direct quantitative evidence of downstream utility on real data.

VNC evaluates real sections without additional simulated deformation; EPFL provides a complementary controlled-perturbation comparison. All methods process the same EPFL deformed input (seed 46), then use the same frozen segmenter, threshold 0.5, and reference annotations. Clean-reference and deformed-input results show performance before and after registration.

Table Q5-2. Complete EPFL segmentation comparison

Method / inputDiceIoU
Clean reference0.803430.67144
Deformed input0.662930.49581
TrakEM20.412660.25997
SEAMLeSS0.692230.52932
EFSR0.446560.28746
NeuroTrajReg0.683750.51947
vEMRec0.747550.59687
ReStack-Mamba0.755170.60664

Dice increases by 0.09224 over the deformed input; Dice and IoU exceed vEMRec by 0.00762 and 0.00977. Together with VNC, this shows that improvements in registration outputs transfer to a fixed segmentation model under both real acquisition conditions and controlled perturbations.

For postprocessing, we test dependence on final filtering by disabling ReStack-Mamba's final Gaussian smoothing, trimmed-mean filtering, and optional displacement-field filtering, with LEA weights and initial-alignment settings fixed. The six volumes are jrc_mus_liver, jrc_mus_pancreas, jrc_mus_liver3, jrc_mus_kidney, jrc_mus_heart, and jrc_mus_skin. The table reports an equally weighted mean across these volumes, giving one aggregate value per metric for each method.

Table Q5-3. Six-volume mean results without final postprocessing

MethodPSNR (dB)SSIMNCCMI
TrakEM221.09160.46110.87250.7912
SEAMLeSS21.17450.46830.87660.8091
EFSR14.55040.33350.39070.2113
vEMRec24.93930.66060.94411.1451
ReStack-Mamba25.17250.70650.95551.1620

Within each volume, PSNR/SSIM are weighted by valid slice counts per view. Archived TrakEM2, SEAMLeSS, and EFSR outputs are evaluated using the common protocol without additional filtering.

Without final filtering, mean PSNR/SSIM remain higher than vEMRec's (25.1725 dB / 0.7065 versus 24.9393 dB / 0.6606); NCC/MI improve by 0.0114/0.0169. Thus, the pipeline retains its image-similarity advantage without final smoothing, showing that this smoothing is not the main source of the observed gains.

Reviewer_C7DvOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ReStack-Mamba, a self-supervised method for aligning stacks of electron microscopy images. It first corrects local distortions, then uses two Mamba stages and a final smoothing step to improve consistency across slices.

Critical strengths
  1. It addresses a practical problem: small alignment errors can disrupt the reconstructed 3D structure.

  2. The design makes sense, combining local alignment with information from three viewing directions without requiring labeled deformation fields.

  3. The full pipeline shows consistent improvements across six simulated volumes, with an additional example on real data.

Critical weaknesses
  1. The contribution of the learned modules is unclear. Table 2 shows a substantial gain from the final smoothing step, while Stage 2 adds very little. How much of the improvement comes from better alignment? It would help to apply the same post-processing to the baselines and compare the results.

  2. The real-data evaluation is limited. The paper only shows qualitative results on one 1,024-slice real-data crop. Testing on more and larger real volumes would strengthen the results. Even without registration labels, downstream tasks such as segmentation could provide quantitative comparisons.

  3. The choice of regularization weights needs more explanation. The paper gives the values but does not explain how they were chosen. How sensitive are the results to these weights, and are the same settings used across all datasets and both stages?

  4. The computational cost is unclear. There are no runtime or memory comparisons with other methods. The paper should report the cost of the full pipeline, since a large increase in inference time could make the accuracy gains less compelling.

  5. The reason for choosing Mamba needs stronger support. Its linear complexity is a reasonable motivation, but the paper does not compare it with other architectures. Comparisons with alternatives such as 3D CNNs or Transformers should report both accuracy and inference cost.

  6. Lack of some related works, such as "Fast and Accurate Electron Microscopy Image Registration" (MICCAI2019), "Electron Microscopy Image Registration Using Correlation Volume" (ISBI2023), and more.

Questions for authors

Please refer to "Critical Weaknesses".

3 comments
AuthorsOpenReview ↗

Response to Reviewer C7Dv (1/3)

W1. Learned refinement versus final smoothing

We thank the reviewer for asking us to separate these contributions. With final filtering disabled, the revised independent Stage 2 outperforms two shared-weight refinement passes, and the complete pipeline retains its advantage over vEMRec.

On EPFL (deformation seed 46), we fix initial alignment, disable final filtering, and use the same frozen segmenter. Values below are means ± sample SD across training initializations 43, 44, and 45. Axial SSIM averages XZ and YZ SSIM.

Table W1-1. Stage comparison without final filtering

ConfigurationTraining epochsAxial SSIMDice
Stage 1240.48364 ± 0.000060.75005 ± 0.00009
Stage 1 twice, shared weights240.48491 ± 0.000100.75213 ± 0.00008
Independent Stage 2 (1×)24 + 240.48473 ± 0.000030.75182 ± 0.00006
Independent Stage 2 (0.5×)24 + 240.48786 ± 0.000020.75517 ± 0.00022

Revised Stage 2 improves Dice by approximately 0.0051 over Stage 1 and by 0.00304 over the shared-weight control; axial SSIM improves by 0.00295 over that control. The original 1× configuration did not outperform shared weights: the measurable increment comes from the revised configuration (full loss and weights in W3, Part 2). Dice provides evidence beyond image similarity.

We evaluated six volumes—jrc_mus_liver, jrc_mus_pancreas, jrc_mus_liver3, jrc_mus_kidney, jrc_mus_heart, and jrc_mus_skin—with ReStack-Mamba’s final Gaussian smoothing, trimmed-mean filtering, and optional displacement-field filtering disabled, while keeping the LEA weights and initial-alignment settings fixed. The table reports PSNR, SSIM, NCC, and MI averaged equally across these six volumes, giving one aggregate value per metric for each method. Within each volume, PSNR and SSIM are weighted by the number of valid slices in each view.

Table W1-2. Six-volume results without final postprocessing

MethodPSNR (dB)SSIMNCCMI
TrakEM221.09160.46110.87250.7912
SEAMLeSS21.17450.46830.87660.8091
EFSR14.55040.33350.39070.2113
vEMRec24.93930.66060.94411.1451
ReStack-Mamba25.17250.70650.95551.1620

The gains over vEMRec are 0.2332 dB, 0.0459, 0.0114, and 0.0169, respectively. Both the learned-stage increment and the full-pipeline advantage remain without final filtering, supporting a contribution from learned registration itself.

W2. Additional real-data segmentation evaluation

We thank the reviewer for suggesting downstream segmentation to quantify real-data performance. Beyond the original Kasthuri11 volume, we added segmentation evaluations on real VNC data without additional simulated deformation and on EPFL data with mitochondrial annotations under controlled simulated deformation. ReStack-Mamba achieves the highest Dice, IoU, precision, and recall among the six methods on VNC, and the highest Dice and IoU on EPFL.

No simulated deformation is added to VNC. All methods use the same frozen segmenter and threshold 0.5; metrics are computed against reference annotations per slice and then averaged.

Table W2-1. Real VNC segmentation (higher is better)

MethodDiceIoUPrecisionRecall
TrakEM20.553690.405540.554840.67112
SEAMLeSS0.663310.496240.619380.71397
EFSR0.574160.408480.560390.68852
NeuroTrajReg0.668320.501860.613600.70706
vEMRec0.664130.497150.623500.71042
ReStack-Mamba0.689210.511650.630100.72828

The respective gains over vEMRec are 0.02508, 0.01450, 0.00660, and 0.01786; gains over the strongest baseline for each metric are 0.02089, 0.00979, 0.00660, and 0.01431. Improvements cover detection, prediction accuracy, and segmentation overlap.

Table W2-2. Complementary EPFL controlled-deformation results

Method / inputDiceIoU
Clean reference0.803430.67144
Deformed input0.662930.49581
TrakEM20.412660.25997
SEAMLeSS0.692230.52932
EFSR0.446560.28746
NeuroTrajReg0.683750.51947
vEMRec0.747550.59687
ReStack-Mamba0.755170.60664

EPFL uses a real image volume with simulated deformation seed 46, the same frozen segmenter, threshold 0.5, and reference annotations. Dice improves by 0.09224 over the deformed input; Dice/IoU exceed vEMRec by 0.00762/0.00977. Together, VNC and EPFL extend qualitative observations to quantitative downstream evidence on real sections and controlled deformations, supporting practical utility for volume reconstruction.

AuthorsOpenReview ↗

Response to Reviewer C7Dv (2/3)

W3. Weight selection, sensitivity, and a common configuration

We thank the reviewer for requesting the rationale and sensitivity of these weights. The initial weights conservatively limit residual changes; we halve all six Stage 2 regularization weights, using the same stage-specific configuration across the main volumes and deformation strengths. EPFL results improve in XY, axial, and segmentation metrics.

We fix initial alignment on EPFL (deformation seed 46), use revised signed NCC in both conditions, and disable final filtering. Only overall Stage 2 regularization strength changes.

Table W3-1. Scaling all six regularizers from 1× to 0.5×

MetricStage 2 (1×)Stage 2 (0.5×)
XY PSNR (dB)20.4584120.48892
XY SSIM0.469700.47260
Axial SSIM0.48473 ± 0.000030.48786 ± 0.00002
Dice, frozen segmenter0.75182 ± 0.000060.75517 ± 0.00022
Fraction of non-positive residual Jacobian determinants01.571 × 10⁻⁷
Mean absolute residual displacement (pixels)0.076410.14616
XY gradient-magnitude ratio to reference0.327890.32832

Values average three training initializations; axial SSIM and Dice include sample SD. Displacement/Jacobian statistics describe the learned residual relative to LEA. A gradient-magnitude ratio closer to 1 is closer to the reference.

Halving regularization increases XY PSNR by approximately 0.0305 dB, XY SSIM by 0.00290, axial SSIM by 0.00313, and Dice by 0.00335, while allowing larger residuals. The conservative initial setting thus restricted useful corrections. Uniform scaling preserves relative weights and increases the influence of similarity.

Table W3-2. Weights shared across the main volumes

Loss termStage 1Stage 2
Adjacent-slice similarity11
Image identity10.5
Displacement magnitude0.50.25
Spatial gradients0.80.4
Boundary and mean-displacement anchoring0.30.15
First-order axial differences0.40.2
Second-order axial differences0.80.4

For stage s∈1,2s\in\\{1,2\\}, IinI_{\mathrm{in}} is the input patch, I^\hat I its refined output, and uu the backward residual field. Stage 2 is independently initialized and trained on fixed Stage 1 outputs. Both stages use:

\mathcal L^{(s)}={}&\mathcal L_{\mathrm{sim}}^{(s)} +\lambda_{\mathrm{id}}^{(s)}\mathcal L_{\mathrm{id}} +\lambda_{\mathrm{mag}}^{(s)}\mathcal L_{\mathrm{mag}} +\lambda_{\mathrm{grad}}^{(s)}\mathcal L_{\mathrm{grad}}\\ &+\lambda_{\mathrm{anchor}}^{(s)}\mathcal L_{\mathrm{anchor}} +\lambda_{z1}^{(s)}\mathcal L_{z1} +\lambda_{z2}^{(s)}\mathcal L_{z2}. \end{aligned}$$ Signed local NCC is evaluated at the original and half scales: $$\mathcal L_{\mathrm{sim}}^{(s)}=-\frac12\sum_{\ell\in\\{1,1/2\\}} \frac{\sum_p w_{\ell p}\rho_{\ell p}}{\max(1,\sum_p w_{\ell p})},$$ $$\rho_{\ell p}=\operatorname{clip}\left( \frac{c_{\ell p}}{\sqrt{(v^A_{\ell p}+\epsilon)(v^B_{\ell p}+\epsilon)}},-1,1\right), \qquad w_{\ell p}=\operatorname{stopgrad}\left(\frac{v^B_{\ell p}}{v^B_{\ell p}+\epsilon}\right), \quad\epsilon=10^{-4}.$$ $A,B$ are the later/earlier refined slices in each adjacent pair; $c$ is local covariance; $v^A,v^B$ are local variances clamped to non-negative values. Sums cover valid $9\times9$ windows over all adjacent pairs. Half-scale images use bilinear resizing with aligned corners. Stop-gradient variance weighting downweights low-texture windows. The six regularizers are: $$\mathcal L_{\mathrm{id}}=\langle(\hat I-I_{\mathrm{in}})^2\rangle, \qquad\mathcal L_{\mathrm{mag}}=\langle u^2\rangle,$$ $$\mathcal L_{\mathrm{grad}}=\frac13\sum_{d\in\\{x,y,z\\}}\langle(\Delta_d u)^2\rangle, \qquad\mathcal L_{\mathrm{anchor}}=\langle u_1^2\rangle+\langle u_D^2\rangle+\langle(\bar u_{zxy})^2\rangle,$$ $$\mathcal L_{z1}=\langle|\Delta_z u|\rangle, \qquad\mathcal L_{z2}=\langle|\Delta_z^2 u|\rangle.$$ $\langle\cdot\rangle$ is an elementwise mean; $\Delta_d$ is a forward difference along axis $d$. $u_1,u_D$ are displacements at the first/last patch slices; $\bar u_{zxy}$ is the spatial mean of each displacement component over the patch. Identity and magnitude terms limit image changes and displacement; gradient/axial terms suppress discontinuity; anchoring constrains boundaries and mean drift. Signed NCC distinguishes positive from negative correlation. Both conditions use this revised similarity, isolating regularization strength. Similarity guides corrections; regularization controls magnitude and continuity. Moderate relaxation yields measurable gains without retuning weights per main volume.
AuthorsOpenReview ↗

Response to Reviewer C7Dv (3/3)

W4. End-to-end computational cost

We thank the reviewer for emphasizing the cost of the complete pipeline. With identical inputs and hardware, ReStack-Mamba reduces mean end-to-end runtime by approximately 7.4% relative to vEMRec, with only 0.027 GiB more peak allocated GPU memory. The accuracy gains therefore do not require longer end-to-end inference in this comparison.

We use the same 100 slices of 800 × 800 pixels on an A100-PCIE-40GB. Each method is warmed up once and timed three times, rotating execution order. Timing includes checkpoint loading and output I/O, but excludes common input preparation and Python startup. Times are means ± sample SD.

Table W4-1. End-to-end cost on identical inputs

MethodTotal time (s)Time/slice (s)Throughput (slices/s)Peak allocated GPU memory (GiB)
vEMRec55.75 ± 0.100.5571.7940.272
ReStack-Mamba51.61 ± 0.440.5161.9370.299

Mean runtime decreases by 4.14 s, and throughput increases from 1.794 to 1.937 slices/s. Thus, the revised method achieves lower mean end-to-end runtime with a small increase in allocated memory. Together with the accuracy advantage without final filtering in W1, these results support a favorable accuracy-cost trade-off for the reported workload.

W5. Mamba versus alternative backbones

We thank the reviewer for suggesting direct backbone comparisons beyond the theoretical linear-complexity argument. Matched backbone comparisons favor Mamba: it achieves higher EPFL Dice and lower mean runtime and peak allocated memory in a separate forward-pass benchmark.

Table W5-1. EPFL segmentation Dice, deformation seed 47 and training initialization 43

BackboneStage 1Stage 2 (0.5×)
Mamba0.752590.75855
3D CNN0.742510.74812
Axial Transformer0.742470.74752

Mamba achieves the highest Dice at both stages. Its Stage 2 Dice exceeds 3D CNN and axial Transformer by 0.01043 and 0.01103, respectively.

For computational cost, we separately use earlier checkpoints that preserve depth resolution and measure two independent stage forward passes on the same patch and RTX A6000.

Table W5-2. Two independent forward passes using the earlier checkpoints

BackboneMean time (ms)Peak allocated GPU memory (MiB)
Mamba51.92721.7
3D CNN59.78723.5
Axial Transformer65.75731.5

These comparisons provide empirical support for choosing Mamba: higher Dice on EPFL and lower time and memory costs for the measured forward workload. The accuracy and cost tables refer to their respective checkpoints and test settings, supporting the architectural trade-off with measurements rather than theoretical complexity alone.

W6. Additional related work

We thank the reviewer for pointing out these missing references. We have added the indicated MICCAI 2019 and ISBI 2023 studies and clarified their technical relationship to our method in context modeling and displacement estimation.

The related-work section now includes the MICCAI 2019 study on electron-microscopy registration using 3D convolutions and the ISBI 2023 study using correlation volumes. The revisions explain how the former uses neighboring-slice context and how the latter combines multiscale matching with recurrent displacement updates. These additions clarify the relationship between our method and existing convolutional context modeling and correlation-based matching approaches.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their time and constructive feedback.

Our point-by-point responses refer to the relevant sections and tables in the revised manuscript, where the changes are highlighted in blue. We hope that our clarifications and additional analyses will resolve the remaining concerns and strengthen the case for acceptance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.