List of Papers Browse by Subject Areas Author List
Abstract
Diffusion bridges provide a principled way to connect paired source and target images for medical image translation, but pixel-space bridge trajectories may still entangle global contrast transport with local texture synthesis. This makes frequency-specific modelling less explicit and can limit the interpretability of cross-contrast MRI translation. To address this issue, we propose \methodfull{} (\method), a frequency-separated paired-translation framework in complex-wavelet space. The proposed method first decomposes each image with a 3-level dual-tree complex wavelet transform (DTCWT). Low-pass coefficients are aligned by a deterministic monotone mapping to obtain a structure-aligned initialization, while modality-specific high-frequency residuals are sampled by Gaussian bridges at each DTCWT level in learned compact texture subspaces. A Dirichlet constraint anchors non-texture regions to the source image, and gradient-isolated multi-stage training reduces shortcut learning while keeping the low-frequency mapping, wavelet-bridge, and image-refinement outputs inspectable. On BraTS 2021, covering T1$\to$T2, T1$\to$FLAIR, and T2$\to$FLAIR translation, \method{} achieved the highest PSNR and SSIM among all baselines, including $27.74{\pm}2.56$ dB PSNR and $92.69{\pm}2.14$\% SSIM for T1$\to$T2. On Gold Atlas, \method{} also achieved the highest PSNR on both MRI-to-CT synthesis tasks. Code is available at https://github.com/xiao9672/RDBWS.git
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3942_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/xiao9672/RDBWS.git
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YanXia_Residual_MICCAI2026,
author = { Yang, Xiao AND Zhao, Wanqing AND Nsengimana, Jérémie AND Zhang, Jingjing},
title = { { Residual Diffusion Bridge in Wavelet Space for Medical Image Translation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
1.The paper proposes Wavelet Residual Bridge (WRB) for paired medical image translation, which decomposes translation into deterministic low-frequency contrast alignment and stochastic high-frequency residual synthesis. 2.It introduces wavelet-subspace Gaussian bridges with a Dirichlet constraint to better restrict synthesis to texture-related regions and preserve non-texture areas. 3.It designs a gradient-isolated multi-stage training strategy to reduce shortcut learning and improve interpretability of intermediate outputs. 4.The method is evaluated on BraTS 2021 and Gold Atlas, with reported improvements over several baseline methods in PSNR/SSIM.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper addresses a clear problem in cross-contrast medical image translation, namely the entanglement of global contrast change and local texture synthesis. 2.The method is technically coherent: the wavelet-space factorization, constrained bridge design, and staged training are well aligned with the claimed motivation. 3.The experiments include multiple baselines, qualitative comparisons, and ablation studies, which make the empirical section reasonably complete. 4.The reported quantitative results are competitive, especially on BraTS where WRB achieves the best PSNR and SSIM across all three tasks.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The validation scale is still limited, with a 40-subject BraTS subset and a 19-patient Gold Atlas dataset. 2.The method is only evaluated in a 2D slice-wise setting, so its value for full 3D medical imaging scenarios remains unclear. 3.The evaluation mainly relies on image-level metrics such as PSNR, SSIM, and MAE, without downstream or clinical validation. 4.Although the pipeline is fairly complex, the paper provides limited discussion of efficiency and practical deployment cost.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
1.I find the main idea interesting and reasonably novel, especially the explicit factorization of structure and texture in wavelet space. 2.The method is not just a minor modification of an existing diffusion model; it includes several components that are conceptually connected and supported by ablation results. 3.The empirical results are strong enough to support that the method is promising, particularly on BraTS. 4.However, the paper is not strong enough for a clear accept because the experiments are still limited in scale and mainly demonstrate improvements on standard image fidelity metrics rather than broader medical utility. Overall, I lean Weak Accept because the paper has a solid methodological contribution and good initial results, but stronger validation would be needed for a higher recommendation.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
They proposed a wavelet-based diffusion bridge for medical image translation task. The authors claim decomposing cross-contrast translation into low and high frequency mappings yields a better separation of roles and improved translation performance.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
-Idea of decomposing translation problem into subparts and approaching them with diffusion bridges is a promising research direction. -Proposed method is clear and technically sound. -Proposed method yields improved performance in widely used public benchmarks against several well-known models.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The authors argue that pixel-space diffusion bridges “entangle global intensity transport with local texture synthesis,” and earlier they say that this entanglement leads to over-smoothed outputs “This makes it difficult to disentangle global contrast adjustment from local detail generation and can produce over-smoothed outputs”. However, we cannot observe over-smoothed outputs from pixel diffusion bridge method SelfRDB[1] neither in qualitative translation figures nor in quantitative tables. Moreover, PSNR and SSIM do not reliably capture over-smoothing; in many cases, especially under uncertainty, they can even favor smoother outputs over sharper ones with small structural deviations. The authors should either provide theoretical justification or direct empirical evidence for this claim. Otherwise, it reads as an unsupported assertion introduced mainly to make the argument sound plausible.
2.The authors also claim improved interpretability through intermediate representations, but they do not actually make meaningful use of these intermediate outputs in the paper. Even if one accepts that the intermediate stages are more interpretable, the practical value of this interpretability remains unclear. The paper should explain what concrete use case this offers.
3.The proposed model includes an image-domain refiner, but the paper provides no motivation for why this component is needed and gives no detail. In its current form, the refiner feels more like an additional engineering component added to push the performance slightly beyond competing methods, rather than a well-justified part of the overall method.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
See major weaknesses. The authors must convince the readers about the motivation behind their approach, their image-domain refiner, and use case of improved interpretability.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes Wavelet Residual Bridge (WRB), a factorized framework for cross-contrast MRI translation that separates deterministic low-frequency contrast alignment from stochastic high-frequency residual synthesis in dual-tree complex wavelet (DTCWT) space. Low-pass coefficients are aligned via a monotone spline map, while high-frequency residuals are modeled as masked Gaussian bridges in learned low-rank subspaces with a Dirichlet constraint anchoring non-texture regions to the source. Gradient-isolated multi-stage training is introduced to prevent shortcut learning. Evaluated on BraTS 2021 (T1 to T2, T1 to FLAIR, T2 to FLAIR) and Gold Atlas (MRI to CT), WRB reports the highest PSNR across all five tasks.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
S1.Well-motivated structure-texture factorization. Decomposing cross-contrast translation into deterministic low-frequency mapping and stochastic high-frequency synthesis is physically sensible. Global contrast differences between MRI sequences are largely captured in low-pass coefficients, while modality-specific texture lives in high-pass subbands. This factorization provides interpretability that pixel-space bridges lack. S2.Thorough ablation study. Table 3 systematically removes each component and quantifies its contribution. The Dirichlet constraint proves critical (removal causes collapse to 15.10 dB), the monotone intensity map has the largest stable impact, and the refiner contributes least. This level of component-wise analysis is informative. S3.Statistical rigor. All metrics are reported as mean +/- std, and paired Wilcoxon signed-rank tests confirm significance. This is a meaningful step above the many synthesis papers that report only mean values. S4.Code availability. An anonymized repository is provided, supporting reproducibility.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
W1.Critically small dataset scale with unclear justification. BraTS 2021 contains 1,251 training subjects, yet only 40 are used with a 5:1:2 split, yielding roughly 25 training and 10 test subjects. Gold Atlas has only 19 patients (approximately 4 test subjects). At this scale, strong quantitative results may reflect overfitting to a narrow distribution rather than genuine generalization. The large standard deviations observed in Gold Atlas results (e.g., SSIM std of 7.40 for BBDM on T1 to CT) are symptomatic of extremely small test sets where a single outlier subject dominates. No justification is given for the extreme subsampling of BraTS, and no cross-validation or repeated-split analysis is provided. This fundamentally limits confidence in all reported results. W2.2D slice-wise processing contradicts the volumetric clinical motivation. All experiments operate on individual 2D axial slices extracted from a fixed index range. Inter-slice consistency, through-plane continuity, and 3D anatomical coherence are entirely unaddressed. The authors acknowledge this as a limitation but do not analyze slice-to-slice discontinuity in generated volumes. For a paper motivated by radiotherapy planning, where 3D volumetric consistency directly affects dose calculation and target delineation, this is a significant gap. W3.No downstream task validation. Evaluation is limited to pixel-level metrics (PSNR, SSIM, MAE). No segmentation, registration, radiotherapy dose calculation, or other task-level evaluation is provided. PSNR/SSIM improvements do not necessarily translate to clinical utility. The disconnect between the clinical motivation (radiotherapy, neuro-oncology) and the purely image-quality evaluation weakens the practical relevance claim.
W4.Framework complexity with marginal gains and missing computational analysis. WRB combines 3-level DTCWT, monotone spline OT mapping, per-level low-rank subspace projections, per-level Gaussian bridge denoisers, Dirichlet constraint masking, and an image-domain UNet refiner, with numerous hyperparameters (subspace dimensions, spline knots, bridge steps, blending schedule, four loss weights, EMA decay). Yet SelfRDB, a single pixel-space bridge, already achieves SSIM above 90% on most tasks. WRB’s improvements are marginal (1-2%p SSIM, 1-1.3 dB PSNR) and metric-dependent: SelfRDB leads in MAE on T1 to FLAIR. No computational cost comparison (training time, inference time, GPU memory) is reported, making it impossible to assess whether the added complexity is justified. Moreover, the method section dedicates substantial space to re-deriving standard formulations (Brownian bridge marginals, wavelet decomposition, linear subspace projection) while the architecture of the core learnable component, the per-level bridge denoiser, is left unspecified.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The choice to use only 40 BraTS subjects deserves explicit justification. If computational cost is the bottleneck, reporting wall-clock training time and scaling experiments on 100, 500, and 1000+ subjects would significantly strengthen the paper. The fixed axial slice index range selection is unexplained and could introduce selection bias. A downstream segmentation experiment using synthesized contrasts as nnU-Net input would directly test clinical utility and is feasible with BraTS annotations. On T1 to FLAIR, SelfRDB achieves lower MAE despite WRB leading in PSNR and SSIM. Given the substantially higher complexity of WRB, the marginal improvements over SelfRDB on certain tasks raise the question of whether the added complexity is justified. A computational cost comparison (inference time, memory, training time) against baselines would help readers assess this trade-off.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
WRB presents a principled structure-texture factorization in wavelet space, supported by a thorough ablation study and statistical testing. These qualities place it above average in methodological rigor. However, the extremely small dataset scale (25 training / 10 test subjects from BraTS, 4 test subjects from Gold Atlas) without justification is a serious concern that undermines confidence in generalizability (W1). The 2D-only design contradicts the radiotherapy motivation (W2), no downstream task evaluation is provided (W3), and the framework’s substantial complexity yields only marginal, metric-dependent gains over the much simpler SelfRDB baseline without any computational cost analysis (W4). The method section also leaves the bridge denoiser architecture unspecified while devoting significant space to standard formulations.If the rebuttal provides a convincing justification for the dataset scale and additional analysis, the score could be reconsidered.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal addressed W3 and W4 satisfactorily. WRB’s inference speed (0.385 s/slice) is nearly identical to SelfRDB (0.372 s/slice), alleviating the complexity concern. The downstream nnU-Net results (WT/TC/ET Dice close to ground-truth inputs) suggest that the outputs are clinically meaningful for the missing-contrast recovery scenario. W1 (dataset scale) and W2 (2D-only evaluation) remain unresolved, but do not outweigh the methodological contribution. Raising my recommendation from Reject to Accept.
Author Feedback
We’d like to thank the reviewers for their positive feedback and constructive comments. The primary concerns are addressed below and will be incorporated into the final manuscript. 1.Clinical use case [R4-W3]: We clarify that WRB targets aired medical image translation for radiotherapy and neuro-oncology applications, including synthesizing FLAIR from T1/T2 MRI scans, and creating CT images from MRI data. These synthesized image contrasts represent key use cases, as they can recover missing scans and reduce examination time and cost by avoiding acquisition of all scans. As the reviewers noted, the work may support downstream task such as segmentation. Although outside the current scope, they will be briefly addressed during downstream validation. 2.Dataset scale/splits [R2-W1/R4-W1]: Regarding the MRI BraTS dataset, we followed previous studies (e.g. SelfRDB[1]) by randomly sampling 4,000 T1-T2-FLAIR paired images from 40 patients, and adopting a training-validation-test split at a ratio of 0.625:0.125:0.25.And, for the Gold Atlas dataset, 1,003 MRI–CT pairs from 19 patients were used with the split ratio at 0.6:0.2:0.2. While both datasets are used, we recognize the need to use a large dataset in future evaluations. To enable fair comparisons, we will release data splits and add patient-level evaluation results. 3.2D setting and slice range [R2-W2/R4-W2]: WRB currently operates in a 2D slice-wise setting. However, its low-pass transport, high-frequency residual bridges, and subspace projection in DTCWT space are not inherently limited to 2D. We will discuss the additional work required to design and validate the 3D variants of the DTCWT and UNet denoisers. We adopted the central-slice protocol, i.e., fixed slice range [27,127), to cover tumor voxels in the BraTS subset. This is consistent with previous studies in SelfRDB[1] and SynDiff[arXiv:2207.08208] 4.Downstream validation [R2-W3/R4-W3]: We conducted an additional segmentation task using BraTS dataset. We use pre-trained nnU-Net[arXiv:1809.10486], taking T1, FLAIR,T1ce and (WRB vs ground-truth) T2 as inputs. This yielded WT/TC/ET Dice scores of 0.886/0.857/0.847 and 0.916/0.854/0.845 using WRB-generated and ground-truth T2.This demonstrates the value of WRB outputs in downstream applications, knowing that ground-truth is often unavailable in clinical practice. 5.Wavelet decomposition and interpretability [R3-W1/W2]: We’d like to clarify the statement on over-smoothed outputs concerns a standard DDPM diffusion (as opposed to a diffusion bridge). To address this issue, WRB uses Wavelet decomposition to model contrast transport and high-frequency textures separately. Table 3 confirms this benefit: the DDPM and WRB achieved PSNR/SSIM of 25.97 dB/90.24% and 27.74 dB/92.69% on the T1→T2.We will also include intermediate high- and low-frequency outputs using energy maps to improve explainability. 6.Image-domain refiner, bridge denoiser, and computational efficiency [R3-W3/R4-W4]: (1) The image-domain refiner is to model remaining residuals (e.g. ringing artifacts) from the reconstructed DTCWT output x^recon. This is realized by a unet to predict the final synthesized image, i.e., x^recon + unet(x^recon, y). The PSNR/SSIM gained with and without the image-domain refiner were 27.42/92.29 and 27.74/92.69, respectively. (2) The forward diffusion involves three DTCWT high-frequency bands(128×128/64×64/32×32),each with a level-specific UNet that predicts the clean subspace state z0 from noised zt. The unet is trained band-by-band in coarse-to-fine order. We will update the framework diagram (Fig. 1) and provide network specifications with clearer notations and learnable parameters, to illustrate how they are used to generate noisy examples that residual U-Net learns to restore. (3) We will report experimental set-up(RTXA800, 20 GB peak memory) with results including inference time from WRB (0.385 s/slice), SelfRDB (0.372 s/slice) and 1000-step DDPM (43.484 s/slice).
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The reviewers highlighted multiple positive aspects to this paper. Wavelet decomposition was seen as a well-motivated approach and the method was highlighted as conceptually well put together. Ablations and statistical testing were received positively. Most concern in reviews came from a lack of extensive validation of the method and the clinical viability of a 2D method. This rebuttal should focus on: 1) justification and explanation of the small training and testing datasets, 2) reporting subject level statistics and showing 3D consistency or a plan for 3D extension, 3) explaining further the need and benefits of the image domain refiner. However, please read all reviews completely to determine which other aspects might be useful to include.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors have provided a convincing and appropriate rebuttal that addresses the reviewers’ concerns. The reviewers highlight the paper’s methodological contribution, which outweighs concerns about limited validation. I recommend acceptance of this paper.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors provided a highly convincing rebuttal that successfully addressed initial reviewers concerns regarding computational complexity and clinical validity.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The main issues raised in the initial review were validation scope: small training/test subsets, a 2D-only pipeline, and no downstream task, plus requests to justify the refiner and the over-smoothing motivation.
The rebuttal clarified the refiner’s role, corrected the over-smoothing wording, and reported inference cost showing the added complexity is essentially free at test time relative to SelfRDB. R4 moved to accept, partly on a downstream segmentation result that is outside rebuttal scope and that I discount; even setting that aside, the method’s intrinsic quality was commended by all three reviewers and I found R2’s accept bases on the in-paper evidence rather than anything new. The dataset-scale concern is only partly resolved and should be flagged or acknowledged for the camera-ready, but the unresolved items are scope limitations rather than flaws in the method. With two reviewers at accept and the methodological contribution well regarded, I lean toward acceptance, while noting the validation-scale caveat.