acceptodds
Withdrawn from ICLR 2027

Learning Reliable Evidence for Coarse-to-Fine Deformable Medical Image Registration

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Medical image registration is fundamental to quantitative anatomical analysis and the assessment of pathological changes. Existing methods commonly use multiscale features to progressively align global structures and local details. However, downsampling can introduce aliasing into coarse-scale representations when high-frequency components are insufficiently attenuated. The resulting changes in local feature responses can disrupt anatomical correspondence estimation. We propose EviGuardMorph, a framework that improves registration reliability through the construction and use of correspondence evidence. The Frequency-Biased Evidence Encoder (FBE) applies scale-dependent Gaussian prefiltering before downsampling to construct structure and detail evidence streams with different effective bandwidths. The detail stream is retained only at high-resolution scales, where it interacts with the structure stream through reversible coupling. The Baseline-Protected Bidirectional Matcher (BPBM) preserves the one-way local matching estimate as a baseline. It combines matching entropy with forward-backward cycle consistency to weight bidirectional candidate residuals, suppressing erroneous updates from ambiguous matches. Experiments demonstrate that EviGuardMorph achieves leading registration accuracy across multiple datasets. Spectral analysis further shows that prefiltering reduces the out-of-band spectral energy fraction before downsampling by 90.4%–99.7%.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical Image Registration, Coarse-to-Fine Learning, Evidence Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GmcjOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EviGuardMorph, a coarse-to-fine deformable medical image registration framework. It aims to reduce frequency aliasing caused by downsampling and prevent the propagation of matching errors.The method contains two main modules. FBE uses scale-dependent Gaussian pre-filtering, dual structure-detail branches, and invertible feature coupling to build multi-scale matching evidence. BPBM uses matching entropy and forward-backward cycle consistency to weight bidirectional matching residuals. It also retains one-way matching as a baseline. The method does not require ground-truth deformation fields for supervision. It is evaluated on four datasets: IXI, LPBA40, ACDC, and AbdomenCT. The results are competitive in terms of Dice, HD95, and ASSD. Overall, the proposed framework shows potential and achieves competitive results. However, the link between frequency suppression and improved registration performance is not sufficiently validated. The ability of BPBM to identify incorrect correspondences also requires further investigation.

Critical strengths
  1. This paper addresses frequency aliasing and incorrect matching in multi-scale image registration by proposing a framework that integrates frequency-aware feature extraction with reliability-guided bidirectional refinement. The overall design is well motivated and demonstrates a reasonable degree of innovation and research value.

  2. The proposed method achieves the best average Dice scores and the lowest average HD95 values across four medical image datasets. The authors also conduct comprehensive experiments, including statistical significance tests, module ablation studies, frequency analysis, and failure case analysis, demonstrating strong registration performance and a certain degree of generalization capability. Furthermore, the method does not require ground-truth deformation fields or segmentation labels during training, which enhances its practical applicability.

Critical weaknesses

1.Although pre-filtering reduces out-of-band spectral energy by 90.4%–99.7%, the resulting Dice improvement is only 0.10 percentage points. The connection between spectral suppression and actual registration performance gains remains insufficiently established. Furthermore, there is insufficient evidence to demonstrate that BPBM offers substantial advantages over simpler confidence-weighting strategies.

2.The paper uses matching entropy and forward–backward cycle errors as indicators of correspondence reliability, but it does not establish that these metrics can accurately identify incorrect matches. The effectiveness of the reliability weights therefore remains insufficiently validated by direct evidence. Meanwhile, the improvements in registration accuracy are accompanied by relatively high deformation field folding rates, reaching 0.364% on IXI and 2.411% on AbdomenCT, both substantially higher than those of some competing methods. This suggests a potential trade-off between anatomical alignment accuracy and deformation plausibility.

Questions for authors
  1. The paper reports a 90.4%–99.7% reduction in out-of-band spectral energy, yet Table 8 shows that fixed pre-filtering yields only an additional 0.10 percentage-point improvement in Dice. How can the discrepancy between spectral suppression and actual registration gains be explained? Does this suggest that spectral attenuation does not effectively improve anatomical correspondence?

  2. The paper constructs reliability weights using local matching entropy and forward–backward cycle errors. However, these two indicators do not necessarily imply that the correspondences are correct. Do the existing experiments demonstrate a clear correlation between the proposed reliability weights and actual registration errors, and show that the weights can effectively identify incorrect matches?

  3. Does BPBM provide substantive improvements beyond a simple confidence-weighting strategy?

1 comment
AuthorsOpenReview ↗

Response to Reviewer Gmcj

Q1. Spectral suppression versus the Dice gain.

Response. The prefilter acts on the stability of the coarse displacement fields, which we now measure directly. We trained the full model and the model without the fixed prefilter with three seeds each and shifted both input images by one voxel. An alias-free pyramid would shift its displacement fields by exactly one voxel, so the remaining deviation measures aliasing. Reduction of this deviation with the prefilter (test pairs showing a reduction in brackets):

FieldACDCLPBA
S5-34% (30/30)-54% (10/10)
S4-40% (30/30)-52% (10/10)
S3-37% (30/30)-37% (10/10)
S2-17% (30/30)-18% (10/10)
Final-14% (30/30)-21% (10/10)

The reduction holds at every scale and in every seed (Wilcoxon p = 1.9 × 10^-9 on ACDC, p = 0.002 on LPBA), and it is largest at S5-S3, where aliasing is strongest. The spectral suppression therefore translates into markedly more shift-stable coarse fields. The prefilter adds 0.10 DSC points on ACDC and 0.11 on LPBA (three-seed means), with the full model higher in every seed. FBE as a whole also transfers to other networks:

ModelACDC DSC ↑ACDC Folding (%) ↓LPBA DSC ↑LPBA Folding (%) ↓
RDP0.79360.00160.72320.0202
RDP + FBE0.80720.00090.72550.0274
VoxelMorph-20.79030.14800.59793.6850
VoxelMorph-2 + FBE0.81040.08200.61152.6930

We now summarize the shift-consistency results in Sec. 4.3 and provide the experimental details and full results in Appendix A.2.

Q2. Whether the reliability weights identify incorrect matches.

Response. We tested this in two ways, with labels that use no entropy, cycle error or R.

Real registration errors. On the test pairs, a location is labelled incorrect when the segmentation warped by the ungated candidate disagrees with the fixed segmentation on more than half of its foreground voxels. The AUROC of 1 - R is 0.740 on ACDC and 0.759 on LPBA at S3, and 0.628 and 0.698 at S2, against 0.50 when R is shuffled. It is higher than entropy alone or cycle error alone at both scales on both datasets.

Regions without a valid correspondence. Test images are warped by a known smooth field, and 32^3 blocks of the fixed image are given no counterpart in the moving image, by shuffling their voxels (Shuffle) or inserting anatomy from another subject (Foreign). AUROC (mean ± SD, three seeds):

CorruptionScaleACDC 1-RACDC R permutedLPBA 1-RLPBA R permuted
ShuffleS50.540 ± 0.0100.487 ± 0.0060.530 ± 0.0260.493 ± 0.006
S40.797 ± 0.0310.500 ± 0.0000.953 ± 0.0060.507 ± 0.006
S30.750 ± 0.0440.500 ± 0.0000.893 ± 0.0150.500 ± 0.000
S20.690 ± 0.0300.500 ± 0.0000.523 ± 0.0380.500 ± 0.000
ForeignS50.547 ± 0.0150.483 ± 0.0060.550 ± 0.0100.493 ± 0.012
S40.793 ± 0.0380.500 ± 0.0000.963 ± 0.0060.500 ± 0.000
S30.697 ± 0.0400.500 ± 0.0000.883 ± 0.0150.500 ± 0.000
S20.567 ± 0.0310.500 ± 0.0000.553 ± 0.0060.500 ± 0.000

1 - R separates invalid from valid locations at S4 and S3 (0.70-0.96), while R randomly permuted across locations stays at chance (0.48-0.51). The response is weaker at S5 and at S2 on LPBA, indicating that the sensitivity of R depends to some extent on the anatomy and the scale.

Q3. BPBM versus simple confidence weighting.

Response. To compare BPBM with simple confidence weighting, we replaced its reliability weight with a fixed blend (R ≡ 1), entropy-only or cycle-only weighting, changing only the S5-S2 update:

VariantACDC DSC ↑ACDC Folding (%) ↓LPBA DSC ↑LPBA Folding (%) ↓
One-way baseline0.81090.02360.72690.0886
Fixed blend (R ≡ 1)0.81230.02450.72540.1082
Entropy weight only0.81110.01560.72660.0848
Cycle weight only0.81150.02430.72650.0974
Full BPBM0.81270.01490.72780.0831

Full BPBM has the highest DSC and the lowest folding on both datasets. On LPBA a fixed blend falls below the one-way baseline in DSC and raises folding, whereas the reliability weight turns the same bidirectional candidate into a gain. Neither entropy nor cycle error alone achieves this.

Folding rates. All methods were trained with the same smoothness-loss weight, not tuned per method, so the folding rates compare the architectures at a common regularization level. Against SACB-Net, the strongest baseline, our DSC advantage comes with lower folding on AbdomenCT (2.41% vs. 3.07%) and ACDC (0.016% vs. 0.070%). RDP and ModeT have lower folding on AbdomenCT (0.34% and 0.56%) but also lower DSC, by 2.07 and 6.81 points. Folding can be traded against DSC through the smoothness weight, and we will discuss this in the revision.

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

The paper proposes a framework for deformable image registration called EviGuardMorph, which combines frenquency biased evidence (FBE) encoder and baseline protected bidirectional matcher (BPBM). FBE is gaussian based filtering that creates ‘evidence’ (though this term evidence was never really introduced and explained, however it reads like to be feature prefilter feature map) that preserves local details when the multiscale features for registration are generated. BPBM in turn works on displacement field (which in the paper is named as the sampling displacement field) which seems to be an average of between forward (fix image to moving image) transformation (named as accumulated displacement) and backward transformation (moving to fix image), which is further enhanced by estimation of ‘reliability’ combining uncertainty and ‘directional agreement (which is never explained).

The novel part claimed as main contribitions are FBE and BPBM and the overall new registration framework.

Evaluation is done on four publicly available data sets which include various medical modalities MRI, CT and different organs (so different types of deformations). Evaluation seems to be pretty extensive and shows the superiority of the presented framework

Critical strengths

None

Critical weaknesses

The paper lacks of critical literature review/introduction about what is the state-of-the-art in biomedical image registration. Currently it lists some registration papers, but doesn’t make attempt to discuss why there is a need for scale-adapted encoding. Moreover the papers used in the evaluation section are not all discussed in the introduction, which raises the concern whether the extensive baseline comparison compares against the most relevant methods.

The paper introduces a number of terms (e.g. evidence, reliability, the sampling displacement field, etc) with little or no explanation what they stand for which makes difficult to follow the paper. Some terms are also confusing – what is the difference between the sampling displacement field (denoted by u) and the estimated displacement field increment (denoted by \nable \phi). alignment of terminology with the conventional terminology in the field of deformable image registration would help a lot

Questions for authors

FBE is called frequency-biased evidence encoder, however there is no explanation where the frequency component comes in. it seems to be Gaussian filter smoothing feature maps. How does the local features/details are preserved with Gaussian smoothing?

The concept of bidirectional candidates weighting forward and backward displacement field contributions seems to resemble a concept of inverse consistency (https://doi.org/10.1109/42.932742., symmetric registration: https://arxiv.org/abs/2003.09514

evaluation would highly benefit from a systematic analysis of the proposed concepts - can FBE added to e.g. voxelmorph approach instead of standard encoder? this would show whether FBE provides improvement in the accuracy of image registration. the same applies to BPBM.

1 comment
AuthorsOpenReview ↗

Response to Reviewer zmvL

W1. Why is scale-adapted encoding needed, and are the baselines relevant and adequately discussed?

Response. We will reorganize Sec. 2 around the three families we compare against and discuss every evaluated baseline there. Classical optimization (SyN) is iterative. Learned deformation prediction (VoxelMorph, CycleMorph, TransMorph, LKU, and the band-limited Fourier-Net, WiNet and LessNet) regresses the field from a single encoder. Multiscale and matching-based methods (Im2Grid, ModeT, GroupMorph, RDP, SACB-Net) estimate the field coarse to fine, so coarse displacements pre-align finer features and coarse errors propagate. This last family shares our design and gives the strongest baselines on every dataset, which is why scale-adapted encoding matters: aliasing in coarse features can introduce displacement errors that propagate to finer scales.

W2. What do evidence, reliability and directional agreement mean, and how do uu, Δφs\Delta\varphi_s and φs\varphi_s differ?

Response. Evidence denotes the encoder features used for explicit matching (Sec. 1, Fig. 2). In Eq. (1), uu is the generic sampling displacement of the warping operator. Δφs\Delta\varphi_s is the displacement increment estimated at scale ss, and φs\varphi_s is the displacement accumulated by composition (Sec. 3.1, Eq. 2). Here, Δ\Delta denotes an increment, not the gradient ∇\nabla. Directional agreement is measured by the forward-backward cycle error (Eq. 12), and reliability Rs∈[0,1]R_s\in[0,1] combines it with matching entropy (Eqs. 11-13). We will collect these definitions in an appendix notation table.

BPBM continuously weights local matching corrections. It retains the one-way estimate bb and scales the residual c−bc-b by λsRs\lambda_s R_s, recovering bb when Rs=0R_s=0.

Q1. Where does frequency bias enter, and how does FBE retain details despite Gaussian smoothing?

Response. Frequency bias comes from scale-dependent Gaussian low-pass filtering before downsampling (Eq. 4), which attenuates frequencies above the next grid’s Nyquist limit. Fixed prefilters act on scale transitions. S1 features are retained before these transitions, and the detail stream continues to S2 with weaker filtering (σ = 0.5 versus 0.8 for the structure stream). Coupling and fusion at S1-S2 retain a pathway for fine-scale cues. Table 8 shows why smoothing alone is insufficient: adding prefiltering to a single-stream encoder lowers LPBA DSC from 0.7235 to 0.7208, whereas full FBE reaches 0.7275. This supports combining bandwidth control with the structure/detail design.

Q2. How does BPBM relate to inverse-consistent and symmetric registration?

Response. BPBM regulates local matching corrections rather than enforcing inverse relationships between image transformations. Christensen and Johnson (2001) constrain forward/reverse maps to be inverses. Mok and Chung (2020) construct symmetric diffeomorphic maps through velocity integration and transformation composition.

BPBM forms c=12(f−W(g,f))c=\frac12(f-\mathcal{W}(g,f)) from local feature-matching displacements f,gf,g and updates Δφs=b+λssg⁡(Rs)(c−b)\Delta\varphi_s=b+\lambda_s\operatorname{sg}(R_s)(c-b). Entropy and local cycle error determine RsR_s, which scales only the residual c−bc-b. Small RsR_s attenuates the correction, and Rs=0R_s=0 recovers the one-way baseline bb. We add no inverse-consistency loss and detach RsR_s during training. Thus, consistency weights local corrections without imposing mutual invertibility on the final transformations.

Q3. Do FBE and BPBM improve registration independently, including in other backbones?

Response. We replaced only the encoder in RDP and VoxelMorph-2, retaining their decoders, losses, training budgets and checkpoint-selection rules:

ModelACDC DSC ↑ACDC Folding (%) ↓LPBA DSC ↑LPBA Folding (%) ↓
RDP0.79360.00160.72320.0202
RDP + FBE0.80720.00090.72550.0274
VoxelMorph-20.79030.14800.59793.6850
VoxelMorph-2 + FBE0.81040.08200.61152.6930

One training seed per configuration. FBE improves DSC in all four comparisons. Folding decreases in three, but increases for RDP on LPBA. These results support FBE’s transfer to both tested backbones.

BPBM also improves performance without FBE. In original Table 4, with the SACB-Net encoder and our shared refinement/composition pipeline, replacing SACB-Net’s local attention matcher with BPBM raises DSC from 0.7039 to 0.7235 on LPBA and from 0.5467 to 0.5817 on AbdomenCT. New within-pipeline controls further isolate spatial weighting: relative to Rs=1R_s=1 with learnable λs\lambda_s, full BPBM gains 0.04/0.24 DSC percentage points on ACDC/LPBA, with lower folding (one run per variant). These controls support BPBM’s contribution, including without FBE.

Reviewer_3riDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EviGuardMorph, a coarse-to-fine framework for deformable medical image registration that aims to improve the reliability of anatomical correspondence estimation. The framework consists of two main components. First, the Frequency-Biased Evidence Encoder (FBE) uses scale-dependent Gaussian prefiltering to reduce aliasing during downsampling, while preserving fine-scale anatomical details through separate structure and detail streams with reversible coupling. Second, the Baseline-Protected Bidirectional Matcher (BPBM) uses matching entropy and forward-backward cycle consistency to guide bidirectional displacement corrections while retaining the one-way matching estimate as a baseline. The framework is trained using image similarity and deformation smoothness losses. Experiments on four public datasets covering brain MRI, cardiac MRI, and abdominal CT report improved registration accuracy over existing methods.

Critical strengths

First, he overall results are encouraging. Tables 1-2 report the highest mean Dice on all four datasets, with improvements in boundary distances as well. The abdominal result is particularly relevant because this setting involves substantial inter-subject deformation. The experiments therefore provide more evidence than a single brain-registration benchmark would.

Second, the appendix makes several useful distinctions. Table 8 separates filter placement from the remaining encoder design, and Appendix A.3 defines exactly how the immediate benefit of a correction is measured while holding the upstream field fixed. Appendix A.5 also acknowledges that low entropy and cycle consistency do not ensure anatomical correctness. These qualifications make the scope of the method easier to assess.

Critical weaknesses

My main reservation concerns the evidence for the reliability weighting itself. Equation (8) uses unnormalized dot products, whereas Eqs. (9)-(14) introduce cosine normalization, a temperature, bidirectional matching and a spatial gate together. Table 4 tests this complete change. Figure 3(c) provides useful within-model evidence that the resulting corrections can improve immediate Dice, but does not show that the reliability weights select corrections better than a fixed blend. A normalization-matched forward-only matcher and the same bidirectional candidate with the spatial reliability weight set to one would address this directly. This distinction is central to the claimed contribution.

Besides that, the registration benefit specifically attributable to prefiltering is modest. In Table 8, removing only fixed prefiltering changes LPBA Dice from 0.7275 to 0.7265 and ACDC Dice from 0.8120 to 0.8110. The authors report these results candidly, but no repeated-training uncertainty is given for differences of this size. The large reduction in out-of-band energy establishes that the filter attenuates those frequencies; it does not, by itself, establish that aliasing reduction explains the main accuracy gain. Most of the gain remains without the fixed prefilter.

Questions for authors

As my first question, with the same cosine similarities and bidirectional candidate, how does BPBM perform when the spatial reliability weight is fixed to one and the scale-wise mixing strength remains learnable? How does a normalization-matched forward-only version perform?

Second, are the small prefiltering gains in Table 8 stable across repeated training runs? Please separate this effect from the remaining structure/detail encoder changes.

Finally, how do the abdominal confidence intervals change when dependence between pairs sharing a subject is respected? Does the Dice advantage persist at a comparable folding rate?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 3riD

Q1. Fixed-blend and forward-only controls.

Response. We retrained both requested controls with the same encoder, refinement head, losses and schedule. B2 keeps the cosine normalization, temperature and bidirectional candidate of the full model but sets R ≡ 1 with λ still learnable. B1 is forward-only with the same normalization and temperature.

VariantACDC DSC ↑ACDC Folding (%) ↓LPBA DSC ↑LPBA Folding (%) ↓
B0 one-way, unnormalized0.81090.02360.72690.0886
B1 forward-only, normalized0.81110.01690.72460.1083
B2 fixed blend (R ≡ 1, learnable λ)0.81230.02450.72540.1082
B3 entropy weight only0.81110.01560.72660.0848
B4 cycle weight only0.81150.02430.72650.0974
B5 full BPBM (Ours)0.81270.01490.72780.0831

One training run per variant, means over test pairs.

The full model improves DSC over B2 by 0.04/0.24 percentage points on ACDC/LPBA and over B1 by 0.16/0.32, with lower folding than both. This supports an incremental benefit of spatial reliability weighting beyond a learned fixed blend.

BPBM regulates bidirectional corrections through Δφ = b + λR(c - b), reverting to the one-way baseline b when R = 0. To probe this response beyond aggregate DSC, we corrupted blocks of fixed images generated with known smooth deformations. Shuffle permutes the intensities within a block and Foreign substitutes transformed content from another image. The corruption masks define the positive labels before inference, independently of entropy, cycle error and R. At S4, the AUROC of 1 - R under Shuffle/Foreign is 0.797/0.793 on ACDC and 0.953/0.963 on LPBA. At S3 it is 0.750/0.697 and 0.893/0.883. As a control, randomly permuting the values of R across locations gives an AUROC of 0.48-0.51 at every scale, i.e. chance level. The response is weaker at S5 and S2, indicating that the sensitivity of R depends to some extent on the anatomy and the scale.

Q2. Stability of the prefiltering gain.

Response. The prefiltering gain is reproducible across training runs. We retrained both models with three seeds (31, 32 / 33):

ModelSeed 31Seed 32Seed 33Mean
Full model0.8124 / 0.72760.8131 / 0.72650.8124 / 0.72560.8126 / 0.7266
w/o fixed prefilter0.8109 / 0.72610.8120 / 0.72550.8119 / 0.72490.8116 / 0.7255

The prefilter raises mean DSC by 0.10 percentage points on ACDC and 0.11 on LPBA, and the full model is higher in every seed on both datasets.

To separate this effect from the structure/detail design, Table 8 of the paper gives ACDC DSC of 0.8035 without FBE, 0.8110 with FBE but no fixed prefilter, and 0.8120 for the full model. Replacing the encoder of RDP and VoxelMorph-2 with FBE raises DSC in all four cases:

ModelACDC DSC ↑ACDC Folding (%) ↓LPBA DSC ↑LPBA Folding (%) ↓
RDP0.79360.00160.72320.0202
RDP + FBE0.80720.00090.72550.0274
VoxelMorph-20.79030.14800.59793.6850
VoxelMorph-2 + FBE0.81040.08200.61152.6930

To measure the effect of the prefilter directly, we shifted both inputs by one voxel. An alias-free pyramid would shift its displacement fields by exactly one voxel. With the prefilter, the shift-induced deviation is 34%-54% lower at S5-S3 and 14%-21% lower in the final field, on every test pair of both datasets. This is direct evidence that the prefilter suppresses aliasing in the coarse displacement fields. We now summarize the shift-consistency results in Sec. 4.3 and provide the experimental details and full results in Appendix A.2.

Q3. Subject-level dependence and folding.

Response. The ranking is unchanged and every interval widens by a factor of about 2.5. The 42 test pairs are all ordered pairs of seven subjects. Each draw resamples the subjects with replacement and uses the sample for both the moving and fixed side, so pair (a, b) gets weight n_a n_b. We use 100,000 draws, shared by all methods.

MethodDSC ↑Pair-level 95% CISubject-level 95% CI
Ours0.5905[0.5799, 0.6008][0.5635, 0.6154]
SACB-Net0.5727[0.5602, 0.5855][0.5437, 0.6058]
RDP0.5699[0.5576, 0.5820][0.5372, 0.6009]
ModeT0.5225[0.5110, 0.5342][0.4968, 0.5544]

All methods were trained with the same smoothness-loss weight, not tuned per method, so the folding rates compare the architectures at a common regularization level. Under this setting, our DSC advantage over SACB-Net comes with lower folding (2.41% vs. 3.07%). RDP and ModeT have lower folding (0.34% and 0.56%) but also lower DSC, by 2.07 and 6.81 points. Folding can be traded against DSC through the smoothness weight, and we will discuss this in the revision. The baselines were re-evaluated for this analysis, so their means differ slightly from Table 2 of the paper.

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

Related papers

Discussion (0)

Sign in to comment.