List of Papers Browse by Subject Areas Author List
Abstract
Motion artifacts remain a major challenge in magnetic resonance imaging (MRI) because they degrade image quality and compromise diagnostic reliability. Recently, the strong generative capacity of diffusion models has driven increasing interest in applying these methods to MRI artifact removal. However, conventional noise-driven diffusion relies on injecting Gaussian noise, which does not match the acquisition-induced corruption produced by motion. It also introduces sampling uncertainty and may hallucinate structures. To address these issues, we propose a noise-free cold diffusion for MRI motion artifact removal trained solely on motion-free image. In this framework, the diffusion process is formulated as a physics-driven degradation operator that incorporates an acquisition-consistent masking strategy and a dirichlet-weighted motion state assignment. This design reflects the mechanism through which motion artifacts arise during acquisition and naturally aligns diffusion timesteps with artifact severity. To support inference across diverse motion patterns and artifact levels, we further introduce severity-aware sampling and adaptive degradation estimation to guide the restoration trajectory and improve generalization. Experiments on two public datasets demonstrate that the proposed method achieves superior performance in MRI motion artifact removal.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3538_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiChu_PhysicsDriven_MICCAI2026,
author = { Li, Chuanpu AND Zhang, Jinlong AND Yang, Yuan AND Jin, Yueming AND Yang, Wei},
title = { { Physics-Driven Cold Diffusion with Severity-Aware Sampling and Adaptive Degradation Estimation for MRI Motion Artifact Removal } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16888},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a noise-free, physics-driven cold diffusion framework for MRI motion artifact removal by introducing a severity-aware sampling and adaptive degradation estimation mechanism. Extensive experiments on two public datasets demonstrate its effectiveness.
- 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.
This paper proposes a physics-driven, noise-free, and deterministic cold diffusion framework to address three critical pain points in MRI motion artifact removal: noise mismatch, hallucination, and data dependency. The method is innovative to some extent.
- 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.
The dataset scale is relatively limited, and the comparison methods are not sufficiently up‑to‑date. The ablation experiments and corresponding analysis are superficial and insufficiently in‑depth. Moreover, the demonstrations of generalization ability and clinical reliability are inadequate and lack systematic validation. The authors have not released their source code, which raises doubts about its reproducibility.
- 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 does not provide sufficient information for reproducibility.
- 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?
The dataset scale is relatively limited, and the comparison methods are not sufficiently up‑to‑date. The ablation experiments and corresponding analysis are superficial and insufficiently in‑depth. Moreover, the demonstrations of generalization ability and clinical reliability are inadequate and lack systematic validation. The authors have not released their source code, which raises doubts about its reproducibility.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
The ablation experiments and analyses remain superficial and insufficient. Although the authors attempt to justify the necessity of individual components and explain hyperparameter settings via clinical priors, systematic and rigorous validations are still lacking, which fails to solidify the effectiveness and indispensability of the core designs. In addition, the generalization ability and clinical reliability of the method are not sufficiently validated. The claimed non-rigid motion modeling and real-world robustness lack targeted mechanistic verification. The evaluation merely depends on conventional PSNR/SSIM metrics, without systematic investigations of failure cases and k-space consistency, rendering the experimental results incomplete and less convincing. Critically, the source code is not released in the rebuttal using an Anonymous link, which does not guarantee that the author will open-source the code.
Review #2
- Please describe the contribution of the paper
The paper proposes a physics-driven cold diffusion framework for MRI motion artifact removal that replaces conventional Gaussian-noise diffusion with a deterministic, acquisition-inspired degradation process.
The key idea is to model motion corruption via a k-space-based degradation operator, incorporating acquisition-consistent masking and Dirichlet-weighted motion state assignment. This aligns diffusion timesteps with artifact severity and enables noise-free, deterministic restoration.
Additionally, the method introduces severity-aware sampling and adaptive degradation estimation, allowing the model to generalize across different motion types and artifact levels during inference.
- 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.
-
Very strong and well-motivated idea Replacing Gaussian noise with a physics-inspired degradation process is highly meaningful for MRI and addresses a key limitation of diffusion-based artifact removal.
-
Good integration of physical priors The formulation explicitly incorporates acquisition properties (k-space masking, motion states), leading to a more realistic degradation model compared to purely data-driven approaches.
-
Deterministic diffusion formulation The cold diffusion setup avoids stochastic sampling and potential hallucinations, which is particularly challenging for image-to-image translations.
-
Handling of motion diversity The framework explicitly considers different motion types (rigid and non-rigid components) via Dirichlet-weighted motion modeling, which is a strong and practically relevant design choice.
-
Strong experimental performance The method shows consistent improvements over prior work in both quantitative metrics (Table 1, p.8) and qualitative results (Fig. 2, p.7).
-
- 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.
- Incomplete discussion of related work The paper should include and discuss additional relevant prior work on learning-based MRI motion correction, e.g.:
Küstner T, Armanious K, Yang J, Yang B, Schick F, Gatidis S. Retrospective correction of motion-affected MR images using deep learning frameworks. Magn Reson Med. 2019 Oct;82(4):1527-1540.doi: 10.1002/mrm.27783. Eichhorn, H. et al. (2024). Physics-Informed Deep Learning for Motion-Corrected Reconstruction of Quantitative Brain MRI. In: Linguraru, M.G., et al. Medical Image Computing and Computer Assisted Intervention – MICCAI 2024.MICCAI 2024.Lecture Notes in Computer Science, vol 15007.
-
Simplified assumption on motion-induced k-space inconsistencies The paper states that motion artifacts arise from k-space inconsistencies (p.1–3), but this is only strictly valid under linearized motion assumptions. In reality, motion has a non-linear relationship with k-space encoding, especially for complex non-rigid motion. This assumption should be clarified and its implications discussed.
-
Limited discussion of motion modeling limitations While the method considers multiple motion types, it remains unclear how well the framework captures complex non-rigid or time-varying motion patterns beyond the simulated transformations.
-
Primarily image-domain evaluation The evaluation focuses on reconstructed image quality (PSNR/SSIM), but lacks deeper analysis of: o k-space consistency o robustness to real-world motions o potential failure cases
- 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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
- Please include and discuss additional relevant prior work (e.g., Küstner et al. 2019; Eichhorn et al. 2024) to better position your contribution.
- Clarify the assumption that motion artifacts correspond to k-space inconsistencies, and discuss its validity for non-linear and non-rigid motion.
- It would be valuable to further analyze the method’s behavior under complex motion scenarios, including failure cases and/or realistic motion.
- Consider adding evaluation metrics beyond PSNR/SSIM, such as frequency-domain consistency, task-based metrics, or blinded expert reading.
- Please consider providing access to the source code.
- 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?
This paper presents a strong and well-motivated contribution by introducing a physics-driven cold diffusion framework tailored to MRI motion artifacts. The replacement of Gaussian noise with an acquisition-consistent degradation process is both conceptually sound and practically relevant.
The method is well-designed, demonstrates strong performance, and addresses an important limitation of current diffusion-based approaches. While some assumptions (e.g., k-space inconsistency modeling) require clarification and related work could be better positioned, these are relatively minor issues.
Overall, this is a high-quality contribution with clear impact potential, warranting acceptance.
- 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 have addressed my comments. In contrast to the author’s response, consistency is still analysed on an image level and not on a k-space level. This however does not impair the proposed method. Sufficient justification and clarification is provided for the limitations and realism of the motion simulation. My points were satisfactorily addressed and I remain with my vote for acceptance.
Review #3
- Please describe the contribution of the paper
The paper proposes a cold-diffusion framework for retrospective MRI motion artifact removal. Rather than using a conventional Gaussian noising process, the authors define the forward diffusion trajectory through a physics-inspired degradation operator intended to mimic motion-induced k-space inconsistencies. The method incorporates an acquisition-consistent masking strategy, Dirichlet-weighted motion-state assignment, and two auxiliary components for severity estimation and degradation composition prediction. Experiments on knee MRI data with real motion artifacts and brain MRI data with simulated artifacts report improved PSNR/SSIM relative to baselines.
- 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 work addresses a clinically relevant problem. Motion artifacts remain common in MRI, and robust retrospective correction methods are of practical interest. The authors correctly note that generic Gaussian diffusion processes are not naturally matched to structured acquisition corruption caused by motion. Another positive aspect is the attempt to make degradation severity explicit.
The experimental section includes comparisons against several categories of prior work, including supervised, unsupervised, and diffusion-based methods. Visual examples are generally supportive, and the ablation study suggests that the proposed components each contribute incremental gains.
- 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.
While the paper presents the method as a conceptual departure from noise-based diffusion, the broader idea of replacing Gaussian corruption with task-specific deterministic operators has already been actively explored in MRI. This includes cold diffusion and more broadly diffusion bridge formulation introduced in recent studies (Fourier-constrained DB IEEE TMI 2026, SelfDB ISBI 2025, SelfRDB MedIA 2025 etc). As written, the manuscript does not sufficiently position itself against prior domain-specific diffusion processes designed around MRI image formation, and this point should be clarified by expanding the survey.
A second concern is the realism of the motion model. The proposed degradation operator relies on synthetic mixtures of motion states, weighted masks, and parameterized transforms. Although motivated by MRI acquisition, true patient motion can be substantially more complex, nonrigid, time-varying, sequence-dependent, and scanner-dependent. It is therefore uncertain how faithfully the simulated forward process captures real motion distributions, particularly outside the evaluated datasets.
The empirical gains, are relatively modest in several comparisons. Improvements of around fractions of a dB in PSNR or small SSIM margins may or may not translate into meaningful diagnostic benefit. No reader study or downstream clinical task analysis is provided to support practical significance. Demonstrations on multi-coil raw complex MRI datasets with real and/or simulated motion would strengthen the claims.
There is also some overstatement regarding deterministic sampling eliminating uncertainty or hallucination risk. Deterministic inference does not by itself guarantee anatomical faithfulness, as restoration networks can still hallucinate plausible but incorrect structures.
Finally, the method introduces several handcrafted components and hyperparameters (mask schedule, severity schedule, Dirichlet weighting, auxiliary heads). It is not entirely clear which parts are essential versus heuristic, and whether comparable gains could be achieved with simpler formulations.
- 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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?
The paper targets an important MRI problem and presents a physics-aware diffusion method with encouraging quantitative results. However, the conceptual novelty has to be clarified, particularly given prior work on non-Gaussian and acquisition-aware diffusion processes in MRI. In addition, validation remains somewhat narrow, and the practical significance of the reported gains is not yet fully established.
- Reviewer confidence
Very confident (4)
- [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
Author Feedback
We thank reviewers for affirming our well-motivated physics-driven idea (all), deterministic formulation (R2), motion handling (R2), and strong performance (all). We address main concerns below.
C: Novelty vs. MRI-specific diffusion variants (R3, Meta) R: We respectfully clarify none of the works cited by R3 target motion artifact removal. FDB and SelfDB target accelerated MRI reconstruction with known undersampling masks. SelfRDB targets multi-modal image translation. To our knowledge, no prior cold diffusion is tailored to motion artifact removal. Our novelty is task-specific: (1) acquisition-consistent masking mirroring Cartesian sampling order; (2) Dirichlet-weighted multi-state motion assignment for motion diversity; (3) severity-aware sampling and adaptive degradation estimation for instantiation. We will cite these works and clarify the positioning.
C: Linearized assumption / motion realism (R2/R3, Meta) R: Methodologically, our model does not rely on a linearized assumption. The transformation in Eq.3 is a generic motion operator supporting non-rigid respiratory transformations (Fig. 1), not restricted to rigid translation/rotation. Different motion states corrupt different k-space regions, capturing time-varying corruption. Composition variability is modeled by the Dirichlet-weighted state assignment in Eq. 8.Empirically, R2’s concern on real-world motion robustness and R3’s on simulation faithfulness are addressed by KMAR-50K, real clinical motion data reflecting the non-rigid, time-varying, sequence-dependent, and scanner-dependent variability, across two scanners, six sequences, and three planes. Our SOTA performance on KMAR-50K is direct evidence that our method is effective on real clinical motion artifacts.
C: Component necessity & ablation (R1/R3, Meta) R: Our components each address a specific challenge. Dirichlet weighting captures motion composition diversity. Severity-aware sampling handles unknown artifact severity. Adaptive degradation estimation infers motion composition. Table 2 addresses R3’s question on simpler formulations: M1 is the simpler baseline, and each added component yields consistent gains. The hyperparameters in Implementation Details are grounded in clinical priors (Refs. [20, 26]) rather than heuristic.
C: Evaluation depth (all, Meta) R: Due to space limit, we focus on PSNR/SSIM, the standard metrics used by all compared methods. (1) Dataset scale & generalization: KMAR-50K provides 1,444 paired scans and 62,506 slices of real clinical motion data, across two scanners, six routine sequences, and three planes. HCP further provides brain MRI evaluation widely adopted in this task. SOTA across this diversity evidences generalization. (2) Gain magnitude & clinical value: against PFAD, we achieve +0.41 dB on KMAR-50K and +1.35 dB on HCP. Prior works (Refs. [11, 21]) show that improvements at this level translate into clinical and downstream benefits. Fig. 2 provides visual evidence of diagnostic-detail preservation, complementing a reader study. (3) Failure modes: although not shown due to space limit, our deterministic trajectory avoids stochastic hallucinations, and severity-aware sampling and adaptive estimation handle varying severity and motion composition, reducing failure risk. (4) K-space consistency: the cycle re-degradation loss in Eq.15 explicitly optimizes it during training, implying the metric remains competitive. (5) Comparison breadth: as in Section 3.2, our baselines cover the major paradigms, with PFAD (AAAI 2025) as the most recent SOTA.
C: Hallucination claim (R3) R: Our claim does not concern anatomical faithfulness, but specifically the stochastic-sampling variability of noise-based diffusion (Ref. [13]), which our deterministic trajectory removes. Faithfulness is further constrained by the physics-driven degradation, k-space preservation, and re-degradation loss.
C: Code / Related work (R1, R2) R: We will release code and cite the works mentioned by R2.
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 paper proposes a noise free, physics driven cold diffusion framework for MRI motion artifact removal, replacing Gaussian diffusion noise with a k space degradation operator and adding severity aware sampling plus adaptive degradation estimation for inference. The reviewers recognize the clinical relevance and promising quantitative performance, but raise several important issues. The main issues are limited conceptual novelty beyond existing cold diffusion and MRI specific diffusion variants, the realism and generalizability of the synthetic motion degradation model compared with real patient motion, and the narrow evaluation focused mainly on PSNR and SSIM without reader studies, k space consistency analysis, or downstream clinical validation. Additional concerns include incomplete discussion of prior work, limited analysis of failure modes, modest empirical gains over recent baselines, and insufficient clarity on the necessity of the handcrafted components. Overall, while the paper addresses an important problem and presents an interesting physics aware formulation, the current validation and positioning are not yet strong enough to support acceptance. The issues raised by the reviewers should be addressed during the rebuttal.
- 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 recommend acceptance. The paper presents a timely and well-motivated physics-driven cold diffusion framework for MRI motion artifact removal, replacing conventional Gaussian noising with an acquisition-inspired degradation process. The core idea is interesting and relevant to MICCAI, and the method incorporates several task-specific components, including acquisition-consistent masking, Dirichlet-weighted motion-state assignment, severity-aware sampling, and adaptive degradation estimation. Reviewer 2 remained strongly supportive and felt that the rebuttal satisfactorily addressed the main concerns regarding motion realism and methodological clarification. While Reviewers 1 and 3 continue to raise valid concerns about ablation depth, k-space-level evaluation, failure-case analysis, and clinical validation, the rebuttal provides reasonable clarification and emphasizes evaluation on a real clinical motion dataset with substantial diversity across scanners, sequences, and planes. Although additional validation would strengthen the work, the contribution is sufficiently qualified for acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
This paper proposes a physics-driven cold diffusion framework for MRI motion artifact removal, aiming to replace Gaussian diffusion noise with a k-space degradation operator and improve restoration through severity-aware sampling and adaptive degradation estimation.
The reviewers recognized the clinical relevance of MRI motion artifact removal and appreciated the physics-aware, deterministic cold diffusion formulation. The method also shows promising quantitative performance on both real and simulated motion datasets.
However, reviewers raised concerns about limited conceptual novelty relative to existing cold diffusion and MRI-specific diffusion variants, the realism and generalizability of the synthetic motion degradation model, and the narrow evaluation mainly based on PSNR/SSIM. Concerns also remain regarding superficial ablation analysis, limited failure-case and k-space consistency evaluation, modest gains over recent baselines, and insufficient evidence for clinical reliability.
Although the rebuttal clarified the intended novelty and emphasized results on KMAR-50K real clinical motion data, it did not fully resolve the concerns about component necessity, motion realism, rigorous validation, and practical clinical significance. Therefore, I recommend rejection.
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.
Based on the rebuttal and reviewers’ responses, I recommend accepting this paper. The authors have addressed each reviewer’s comments, particularly regarding novelty and motion modeling, which were raised in multiple reviews. In my opinion, the remaining concerns fall outside the scope of a MICCAI article, given the justifications presented in the rebuttal.