List of Papers Browse by Subject Areas Author List
Abstract
Single-shot echo-planar prostate diffusion-weighted imaging (DWI) is frequently complicated by geometric distortions, which impact the ability to derive reliable diagnoses from such images. Developing automated correction methods is challenged by the absence of paired distorted and undistorted clinical scans. In this paper, we first propose a novel weakly-supervised image quality transfer (IQT) framework from undistorted to distorted images that utilizes image quality assessment (IQA) signals to supervise the transfer process. Unlike traditional methods that require expensive, voxel-wise paired data or resort to developing unpaired algorithms, our approach utilizes image-level quality labels (here, distorted vs. undistorted) to establish latent quality prototypes within a pre-trained feature space. Recognizing that simulating realistic distortions is more reliable than direct unpaired correction, we describe a weakly-supervised prototype flow matching algorithm to explicitly regularize generative trajectories towards distorted prototypes, producing realistic susceptibility artifacts that mimic clinical degradations. By synthesizing these realistic pairs, we enable a second IQT model to be trained in the forward direction for distortion correction. Experimental results demonstrate that our generated images successfully mimic the diagnostic interference of real-world artifacts, which leads to more capable distortion correction IQT models. In addition to qualitative comparisons, we also conduct exhaustive quantitative evaluations that compare our approach with existing unpaired approaches (e.g., CycleGAN, UNIT-DDPM, and OT-FM) - as either forward or reverse alternatives - by assessing clinical downstream task performance in PI-RADS and Gleason score classification, using both in-distribution and external data sets. Code is available at https://github.com/yucheng722/ProtoFM-IQT.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4009_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/yucheng722/ProtoFM-IQT
Link to the Dataset(s)
N/A
BibTex
@InProceedings{TanYuc_Learning_MICCAI2026,
author = { Tang, Yucheng AND Yan, Wen AND Ng, Alexander AND Thorley, Natasha AND Rajwa, Pawel AND Wang, Yipei AND Asif, Aqua AND Allen, Clare AND Dickinson, Louise AND Giganti, Francesco AND Atkinson, David AND Punwani, Shonit AND Alexander, Daniel C. AND Saeed, Shaheer U. AND Kasivisvanathan, Veeru AND Hu, Yipeng},
title = { { Learning to Distort: Weakly-Supervised Image Quality Transfer for Prostate DWI Correction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a two-stage weakly-supervised IQT framework for prostate DWI distortion correction. Instead of direct unpaired correction, the method first synthesizes realistic distortions from undistorted inputs using prototype-guided flow matching with image-level quality labels, then trains a supervised U-Net on the resulting synthetic pairs for correction. The key technical component is a prototype guidance (PG) loss that steers the FM generation trajectory toward the distorted feature manifold using learned quality prototypes.
- 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 idea that distortion simulation is safer than direct correction is intuitive and reasonable. While the synthesis-first strategy itself is not entirely new in medical imaging, the specific combination of defining the target domain via IQA-learned prototypes and injecting it as a weak, time-weighted supervision on the FM trajectory is, to my knowledge, an original formulation.
2.The PG loss provides continuous directional force toward the distorted cluster center, unlike BCE-based guidance where gradient diminishes once the decision boundary is crossed. The ablation study clearly validates this design choice.
- 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.Correction performance is not convincing enough. The gap is substantial, especially for Gleason task where texture fidelity matters. Since the paper’s ultimate goal is correction, this limited performance weakens the practical value of the entire framework. The paper itself notes blur and loss of fine-grained detail. This makes the choice of an L1-only correction objective questionable, although the current evidence is not enough to prove causality.
2.Data labeling and sample size concerns. The distorted/undistorted labeling criterion is described only as “geometric consistency between T2WI and DWI” without specifying the number of raters, inter-rater agreement, or quantitative threshold. The class imbalance (790 vs. 237) and its effect on contrastive learning and prototype learning are not discussed.
- Please rate the clarity and organization of this paper
Good
- 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
1.The L1 objective in L_corr is stated to be chosen ``to preserve high-frequency textural details,’’ yet the paper itself reports the opposite outcome (over-smoothing, degraded Gleason ACC). Could the authors discuss why this loss was chosen and whether alternative objectives were considered?
2.It is unclear whether ADC was recomputed after DWI correction during downstream evaluation. If the original ADC was retained, this modality mismatch could confound the interpretation of correction performance.
- 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?
The paper reframes DWI distortion correction as a ‘learn-to-distort-then-invert’ problem, and the choice of guiding the FM trajectory in IQA feature space is technically sound. However, the final correction performance shows substantial gap from undistorted reference, and statistical rigor is limited given the small sample size. Overall, the paper presents an interesting idea with reasonable execution, but the experimental evidence does not fully support the claims. Rebuttal addressing correction performance gap, and data discrepancy could tip the balance toward acceptance.
- 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.
My concern was partially resolved. The rebuttal clarified the labeling process, including expert-based visual assessment of DWI-T2 gland boundary deviation. This was helpful for my concern about vague distortion labels.
Review #2
- Please describe the contribution of the paper
The author proposed a two-stage weakly supervised framework for prostate DWI distortion correction. In the first stage, they used a prototype-guided flow matching model to synthesize distorted DWI from undistorted scans under weak image-quality supervision. In the second stage, the synthesized distorted–undistorted pairs are used to train a supervised U-Net for distortion correction. The method is compared with several unpaired image translation baselines, including CycleGAN, UNIT-DDPM, and OT-FM, and is evaluated through qualitative comparisons as well as downstream PI-RADS and Gleason score classification on internal and external prostate MRI datasets.
- 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.
The paper is well-structured and clearly written, presenting the method logically with sufficient implementation detail to understand it. The proposed two-stage formulation is conceptually coherent: it first reformulates unpaired distortion correction as distortion synthesis, then applies supervised correction using the resulting synthetic paired data. The weakly supervised synthesis strategy—constraining the generative process with quality-aware prototypes in a learned feature space—is an interesting and relatively novel idea in this application setting.
- 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 realism and subject-consistency of the synthesized distortion are not convincingly established. The proposed framework critically depends on the assumption that the first-stage synthesis produces realistic prostate distortion while preserving the underlying subject anatomy. However, the FM supervision is constructed from minibatch OT-coupled source-target pairs with image-difference velocity targets, rather than subject-specific physical deformation or distortion fields. In this setting, it remains unclear whether the model is truly learning a realistic clean-to-distorted transformation for the same prostate, or instead performing cross-subject translation toward a distorted-looking target. Although the prototype-guidance constraint encourages movement toward the distorted feature manifold, it does not explicitly enforce anatomical fidelity. As a result, the paper does not yet convincingly rule out anatomical mismatch or hallucinated distortion patterns. 2.The comparison is not aligned with the distortion-correction task. While the paper is motivated as a prostate DWI distortion-correction method, the main comparison in Table 1 is primarily against unpaired image-translation or generative baselines such as CycleGAN, UNIT-DDPM, and OT-FM, rather than methods explicitly designed for distortion correction. Even if reverse phase-encoding or field-map-based references are unavailable in the target setting, the evaluation would still benefit from stronger comparison against correction-oriented alternatives, such as structure-guided or registration-based single-phase-encoding approaches, or from more direct geometric validation. In its current form, the comparison mainly shows that the proposed method produces images that affect a downstream classifier more strongly than these baselines, but it does not convincingly establish superior distortion correction. 3.Overclaim. The manuscript repeatedly claims that the synthesized images are “clinically realistic” and that the correction model is effective in real-world scenarios, but the supporting evidence is largely indirect. Most quantitative validation relies on performance changes in a single downstream model after replacing the original DWI with synthesized distorted or corrected images. This does not directly demonstrate that the generated distortion is geometrically realistic or that the corrected outputs are anatomically faithful, since the downstream model may also be sensitive to blur, texture, or intensity-distribution shifts rather than true distortion patterns. This concern is reinforced by the paper’s own qualitative observations that “direct forward generation may fill geometric voids with unrealistic intensities, and that the second-stage U-Net tends to produce over-smoothed, blurrier outputs”. More direct evaluation of geometry and anatomy, such as T2 alignment, or overlap-based assessment (Dice) where annotations are available, would be needed to support the current level of claim.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My score is mainly driven by the fact that the paper’s core premise is insufficiently validated. The method depends on realistic distortion synthesis, but the current experiments do not convincingly show that the synthesized images preserve subject anatomy and reflect true distortion formation rather than cross-subject translation toward a distorted domain. This concern is further amplified because most conclusions are drawn from a single downstream evaluation pipeline based on ProFound, which is not sufficient on its own to substantiate such strong claims. In addition, the evidence for effective correction is largely indirect, relying on one downstream model instead of direct geometric or anatomical evaluation. Given that the paper itself reports unrealistic intensity filling in forward generation and blurrier outputs in the second-stage correction, I do not think the current results are strong enough to support the claims of clinically realistic synthesis and effective distortion correction.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
After reading the authors’ rebuttal, I maintain my recommendation of Reject (score 2). I appreciate the clarification, but the rebuttal does not resolve my main concern: the manuscript’s central premise depends on subject-consistent and anatomically faithful distortion synthesis, whereas the current evidence remains indirect.
First, on anatomical fidelity, the authors argue that the hallucination risk in the first stage is less important because the second-stage correction model is trained to map low-quality inputs to real undistorted targets. I do not find this argument sufficient. The second-stage U-Net is trained on synthetic distorted inputs paired with the original undistorted images; therefore, if the synthetic distorted images contain cross-subject translation effects or non-subject-specific anatomical changes, the resulting supervised correction mapping may itself be biased or ill-posed. Prototype guidance toward a distorted feature manifold may encourage domain-level degradation, but it does not by itself establish preservation of subject-specific prostate anatomy.
Second, on realism, the rebuttal redefines “realism” as downstream diagnostic impact. This is a narrower operational definition, but it does not address the original concern that the downstream model may respond to blur, intensity shifts, texture changes, or other distributional artifacts rather than geometric distortion specifically. Thus, downstream degradation or recovery alone does not demonstrate that the synthesized distortion is anatomically or geometrically realistic.
Third, the concern about the evaluation backbone remains unaddressed. The quantitative conclusions rely heavily on a single ProFound-based downstream evaluation pipeline. In the manuscript, ProFound is not established as a peer-reviewed or clinically validated evaluation instrument for this purpose, and the reported reference performance on undistorted images is modest. This makes it difficult to interpret relatively moderate performance changes as strong evidence of clinically realistic synthesis or effective distortion correction.
I understand the authors’ point that additional acquisition-based or registration-based baselines may not fit their retrospective unpaired-data setting. However, this does not eliminate the evidentiary gap: the paper is framed as distortion correction, but the support for actual geometric/anatomical correction remains indirect. The paper itself also reports that direct forward generation can fill geometric voids with unrealistic intensities and that the second-stage U-Net produces blurrier outputs, which further weakens the current claim of clinically realistic correction.
For these reasons, I do not think the rebuttal sufficiently addresses the core validation concern. Therefore, I maintain my score of 2 (Reject).
Review #3
- Please describe the contribution of the paper
The authors proposed a novel DWI distortion correction scheme, which is mainly based on two models; distorted image synthesis model and distortion correction model trained by the synthesized image pairs (distortion-free and distorted). It is based on weakly-unsupervised IQT scheme including IQA encoder and prototype learning and training. Synthetic distortion is controlled by flow matching model so that various levels of distortion can be generated. The authors validated its performance by comparison among related ML-based methods such as C-GAN and proved the superiority of their method quantitatively in diagnostic measures (prediction performance by PI-RADS and Gleason score) and qualitatively (visual assessment).
- 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.
-
customized distortion correction model trained by synthetic data and training data synthesis For the issues of lack of training data both in quantity and quality, they showed clear solution. Also, not just applying existing model architectures, they customized the model including loss functions.
-
sufficient validation experiment and obtained certain achievement Generally, it is quite hard to validate distortion correction itself. Instead they indirectly evaluate the performance by diagnostic measures, i.e. how the correction process contributed to classification performance with not small quantity of 483 cases (trained by over-1000 cases). The proposed method outperformed the competitors of ML-based methods.
-
clear description with details To clarify their complicated methods, certain efforts can be found to make readers understand. Also, they provides the codes.
-
- 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.
- How did you categorize distorted and undistorted images by “geometric consistency”? Quantitatively or visually? If quantitatively, what was the threshold? Those can influence the performance of the model. Please describe and/or discuss if needed.
- Please rate the clarity and organization of this paper
Good
- 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
- “IQA” could be also spelled out as IQT for readers.
- Description of quantity of cases could be improved concisely, for example below. (227 ≥ 4, 256 < 4) -> (227 for score ≥ 4, 256 for < 4)
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Technical soundness, clear description, and certain achievement are worth accepting as MICCAI paper.
- Reviewer confidence
Very confident (4)
- [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 authors appropriately answered the questions and explained against the concerns by the reviewers.
Author Feedback
We thank the reviewers and AC for the insightful comments.
Anatomical Fidelity and Generative Risk (R2):
R2 questioned if OT-FM synthesis introduces hallucination rather than preserving anatomy. We agree that biophysically accurate or subject-specific deformation modeling is important. That is why we adopted a prototype-guided objective that steers trajectories toward a feature-level distortion manifold, to mitigate anatomically implausible synthesis. This promotes artifact-like variations consistent with clinically observed DWI degradation patterns, rather than identity-preserving reconstruction.
Importantly, we suspect that it is precisely the reason why our proposed strategy - first HQ-to-LQ synthesis followed by paired LQ-to-HQ generation - outperformed the alternative forward generation (direct LQ-to-HQ). In this strategy, those hallucinated LQ images have a lower chance to be used in generating HQ images, as input LQ images are always supervised by (thus map to, during inference) real (undistorted) HQ samples. Therefore, in the proposed reverse-synthesis-then-correction approach, the hallucination issue in the first stage is arguably much less important, and sensitivity of the IQT performance to which is greatly reduced.
Overclaiming Realism and Downstream Performance Improvement (R2):
A related issue was raised by R2 regarding the term “realistic”. We agree that this term can be vague and broad, and we would like to carefully clarify our intention. In this work, we define “realism” as reproducing clinically relevant quality that of sufficient diagnostic value, thus measured by our downstream task performance – PIRADS and Gleason scoring accuracy.
Quantitatively, the performances on PIRADS and Gleason are indeed improved consistently (“distortion correction” block in Table 1), although some improvements may be considered moderate, as correctly pointed out by R2 and AC. A potential confusion may come from the comparison study (1st block in Table 1), only for testing the reverse HQ-to-LQ generative step (thus the lower the better), used in our two-stage strategy – we will clarify this point.
We also agree that the R2-suggested additional geometric metrics (e.g., T2 alignment) would quantify the characteristics of the synthesized images. However, we respectively argue that 1) it is highly challenging to annotate pathological regions in both T2 and diffusion images due to their complimentary nature [22], whilst the anatomical overlap (e.g. gland, zone) is considered of less value in cancer imaging; 2) even for the subset cases where such overlap can be measured, its relevance to cancer visibility and detectability can be questioned (and potentially misleading), in both sufficiency and necessity (e.g. local morphology and texture may matter). We will add this brief discussion.
Additional Comparison Baselines (R2):
R2 requested comparisons with structure-guided or registration-based single-phase-encoding baselines, which require additional data acquisition or nontrivial annotation beyond standard retrospective mpMR data. Whilst we appreciate the suggestions, our work instead focuses on unpaired learning under routine clinical data, reflecting real-world development and deployment scenarios. We therefore compare against unpaired generative baselines under the same constraints.
Additional Details (R1, R3):
Radiology labels were assigned independently under routine clinical conditions by 37 clinical experts (5–20y exp). Distortion was labelled when gland boundaries showed visual deviation between DW and T2.We note that assessing distortion remains an open problem, while clinical relevance is supported by downstream improvements. We observed that label imbalance mainly affects prototype frequency rather than pairwise supervision quality, owing to relatively sufficient negative samples in the large training set. We will clarify these details.
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.
This paper proposes a weakly supervised two-stage framework for prostate DWI distortion correction that first synthesizes distorted images from undistorted scans using prototype-guided flow matching, and then uses these synthetic pairs to train a supervised distortion correction model. The reviewers found the idea novel and clinically relevant, but raised several concerns. The most important concern is whether the first-stage synthesis preserves subject-specific anatomy or instead performs cross-subject translation toward a distorted-looking target, which challenges the validity of the synthetic paired data used for correction. Reviewers also noted that the correction performance remains modest and substantially below the undistorted reference, particularly for Gleason classification where texture fidelity is critical, limiting the practical impact of the method. Additional concerns include insufficient detail on the distorted versus undistorted labeling protocol, potential effects of class imbalance, and the lack of direct geometric validation or comparison against distortion-correction specific baseline techniques. Overall, while the framework is promising, the current evidence does not yet fully support the strength of the claims. A rebuttal should directly address the anatomical fidelity of the synthesized distortions, assess how realistic in the synthesized distortions in terms of MRI physics, clarify the labeling methodology, justify the benchmark choices.
- 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.
After rebuttal, I think the authors have addressed several important clarification points, particularly the distortion labeling process. They explained that labels were assigned by 37 clinical experts based on visual DWI-T2 gland boundary deviation, which helps address concerns about label definition and reliability. They also clarified that their goal is weakly supervised correction using routine retrospective data, which partly justifies comparison against unpaired generative baselines rather than acquisition-dependent correction methods. However, Reviewer 2 raises an important remaining weakness that is not fully resolved. The paper’s central claim depends on the synthesized distorted images being anatomically faithful and subject-consistent, but the rebuttal still relies largely on indirect downstream PI-RADS and Gleason performance rather than direct geometric or anatomical validation. The concern that the model may learn domain-level degradation, blur, or intensity shifts rather than true subject-specific distortion correction remains valid. Despite this limitation, the paper presents an interesting and clinically relevant formulation, reasonable execution, and positive support from two reviewers. I therefore recommend acceptance, while noting that the claims around realistic distortion synthesis and anatomical correction should be moderated.
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 major concerns were partially resolved, particularly regarding the remaining data-labeling clarification and the performance evaluation.
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.
Reviewers agree that the paper is innovative, but one reviewer has experimental qualms. While I agree with these issues, I believe that the innovation should carry for the accept/reject decision, with the stipulation that this paper is closer to the margin than others accepted.
