List of Papers Browse by Subject Areas Author List
Abstract
Accelerated MRI reconstruction has garnered increasing attention due to its significant clinical value. Recently, the exceptional capabilities of diffusion models in image generation have led to their widespread application in accelerated MRI reconstruction. However, the inherent noisy diffusion process in these models introduces uncertainty during the reverse diffusion restoration, which can compromise the consistency of the results. Moreover, adding Gaussian noise contradicts the actual MRI imaging process. To address these issues, we propose FilterDiff, a noise-free frequency-domain diffusion framework. In FilterDiff, the diffusion process is modeled as a filtering operation, similar to the MRI acquisition process, thereby eliminating the dependence on noise and simplifying the diffusion procedure. To better capture frequency-domain long-range information, we proposed a Swin-DiTs network, which modifies the DiT transformer network by replacing the self-attention mechanism with Swin-attention to reduce computational cost, and removing the position embedding to mitigate feature artifacts. Extensive experiments on two public datasets demonstrate that our model achieves state-of-the-art performance in accelerated MRI reconstruction, both in in-distribution and out-of-distribution scenarios.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1256_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{SonTao_FilterDiff_MICCAI2025,
author = { Song, Tao and Nie, Fang and Guo, Yi and Xu, Feng and Zhang, Shaoting},
title = { { FilterDiff: Noise-free Frequency-domain Diffusion Models for Accelerated MRI Reconstruction } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15975},
month = {September},
page = {205 -- 215}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper claims to introduce frequency-domain degradation operators to accelerated MRI reconstruction.
- 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.
- Results are reported on two datasets.
- Writing and organization is generally clear.
- Comparisons against a relatively large number of baselines are reported with notable improvements.
- 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 primary motivation for the study is stated as: “Although ColdDiff[2] is a generalized diffusion method in the image domain, the exploration of frequency-domain degradation operators in accelerated MRI reconstruction tasks has not been conducted.” This claim does not appear to be true, Fourier domain degradations are inherent to diffusion bridge methods that use a diffusion process between undersampled and fully-sampled data.
-
The differences of the diffusion process and the degradation operator from prior work on cold diffusion/diffusion bridges for MRI reconstruction are not discussed, weakening claims on contribution/novelty.
-
Experimental results show at times dramatically poor performances from some of the baselines (15-20 dB-% lower quantitative metrics), it would be helpful to check their implementation for any potential suboptimalities.
-
- 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.
- 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/2025/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 — could be 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 tackles an important research problem with a sound approach. However, the literature survey and associated discussions in the paper neglect prior methods’ contributions to degradation operators in diffusion models. Since this is a key contribution claimed, the authors should clarify this point.
- 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.
In their rebuttal, the authors claim that their proposed model, FilterDiff, differs from previous cold diffusion and diffusion bridge methods in that it uniquely uses a k-space diffusion process based on a low-level degradation endpoint (difference from CDiffMR), and it uses only k-space degradations in a non-noise dependent diffusion process (difference from FDB). The latter claim is untrue in that FDB uses an exclusively non-noise diffusion process based on k-space degradations and ends up at a low-level degradation point, thus the novelty claims based on these items are unwarranted. Given that the authors re-iterate on their novelty claims without acknowledging prior art in an accurate manner, I cannot recommend publication of the work.
Review #2
- Please describe the contribution of the paper
This paper presents a method for MRI reconstruction based on cold diffusion. By changing the degradation process in diffusion models from noise addition to k-space undersampling, and training a restoration network to gradually predict the missing k-space data at each step, the method is better fitted with the MRI imaging process. Experiments on fastMRI and IXI datasets show good performance of the proposed approach compared to other diffusion models.
- 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 relatively well-written, with clearly explained ideas and well-organized contents.
- 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 idea in this paper is very similar to a previous paper published in MICCAI 2023, which is not cited or compared. Therefore, the novelty of this paper is questionable and the literature is not adequately referenced. Please see details below.
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
The idea in this paper seems highly correlated with the paper “CDiffMR: Can We Replace the Gaussian Noise with K-Space Undersampling for Fast MRI?”, published in MICCAI 2023 (https://conferences.miccai.org/2023/papers/107-Paper0204.html). The similarities are:
- Both use the cold diffusion model for MRI reconstruction.
- The degradation operation in “CDiffMR” is also k-space undersampling. This paper did not cite “CDiffMR”, which is a highly related work. Due to the similarity between the two papers, I believe the authors should do the following before their paper can be published:
- Clearly explain the differences in methodology and the novelty of their approach
- Compare with CDiffMR in experiments.
Other questions:
The time index seems inconsistent throughout the paper.
- Fig. 1, first panel: Should t and t-1 columns be switched? The current order M_0, M_t, M_{t-1}, M_T seems wrong. Similar comment for Fig. 2.
- Eq. 9: Should t and t-1 be switched? Based on Algorithm 2, the index t goes from T to 1, i.e. decreases through the algorithm.
Section 4: Table 2: What’s the difference between Linear, Sparse, and Dense? Please clarify.
- 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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Novelty is unclear due to inadequate referencing and comparison to closely related previous work.
- 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 authors have addressed my questions. I would suggest including the comparison with CDiffMR in the final version.
Review #3
- Please describe the contribution of the paper
This study proposes the FilterDiff method for accelerated MRI reconstruction, which models the diffusion process as a filtering operation. Experiments conducted on two public datasets demonstrate that the proposed method outperforms existing approaches on both in-distribution and out-of-distribution data.
- 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 work presents an interesting idea that leverages a central sampling mask as a filter to construct the diffusion process.
- Unlike traditional diffusion models that rely on noise addition, the proposed FilterDiff eliminates the introduction of uncertainty during the reverse process.
- Experimental results show that FilterDiff achieves relatively strong performance compared to other diffusion-based methods.
- 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.
-
Many implementation details are not clearly described. For instance, the equations do not include coil sensitivity map estimation, suggesting the use of single-coil data. It remains unclear how the method performs on multi-coil data.
-
Additionally, the manuscript lacks theoretical justification for the proposed approach.
-
- 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 provide sufficient information for reproducibility.
- 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/2025/en/REVIEWER-GUIDELINES.html
-
In Algorithm 2, it is stated that k_T can be computed from k_c and M_T. Is it a mistake? It should be that k_c can be computed from k_T and M_T.
-
Additionally, the choice of network architecture and sampling parameters in diffusion models can significantly influence performance. Since the compared methods may use different networks, I recommend that the authors separate the performance gains attributable to the proposed method itself from those arising due to network design.
-
- 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?
FilterDiff introduces a deterministic diffusion process, which is both innovative and effective. The experimental results consistently show superior performance, supporting the method’s practical value.
- 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
Author Feedback
We propose a k-space diffusion-based MRI reconstruction method that models the diffusion process between under-sampled (low-degraded) and fully-sampled k-space data. Additionally, we introduce a novel k-space loss function and recovery process tailored to the k-space (R1,R2,R3). We sincerely thank the reviewers for their valuable feedback.
Differences from prior work & Fourier domain degradations (R1): Previous diffusion models have progressed from conditional to locally conditional frameworks, and further to diffusion bridges, with increasingly deterministic final states. Similarly, degradation operators have evolved from monotonic noise addition to bidirectional scheduling, k-space undersampling (CDiffMR [3]), and hybrid noise–sampling schemes (FDB [1]). Building upon these advances, our method introduces a novel, noise-free diffusion process directly in the k-space domain. Unlike CDiffMR [3], which defines the final state as a high-degraded k-space data, our approach uses a deterministically low-degraded k-space data. Moreover, we design novel training losses and a recovery process in k-space, rather than directly adopting ColdDiff’s image-domain losses and recovery process as done in CDiffMR.
Existing works on Fourier Diffusion Models differ from ours: they replace the scalar forward process with shift-invariant convolution and use additive stationary noise instead of white noise.
Comparison with CDiffMR (R2): We appreciate the reviewer’s reminder. While we compared with CDiffMR’s extended version FDB[1], we neglected to cite and directly evaluate CDiffMR[3] itself—this will be corrected in the final version. As shown in FDB Tab. 1 and 2, CDiffMR performs worse than FDB, while our method outperforms FDB, indirectly supporting our superior performance over CDiffMR. Although CDiffMR shares a similar motivation, our work introduces key novelties: (1) The final state M_T in CDiffMR’s diffusion is a high-degraded (near zero) k-space, rather than low-degraded (under-sampled) k-space , making the added diffusion from low-degraded k-space to high-degraded k-space redundant and potentially detrimental to training. (2)While CDiffMR operates in the k-space, it largely inherits ColdDiff method. Our method introduces novel loss (Eq. 8) and recovery process (Eq. 9) specifically designed in the k-space, and additionally use the k_c as a condition. (3) CDiffMR uses 1000 diffusion steps, and although data consistency reduces sampling steps to 35–55%, hundreds of iterations are still needed. Our method completes both degradation and reconstruction in only 20 steps, offering significantly higher efficiency.
Performance Gains from Network Design (R3): As shown in Tab. 2 (Rows 4 and 6), performance improvements stem more from the proposed method than from specific network architecture modifications.
Baseline Performances (R1): Baselines were reproduced using publicly available code, with some results directly adopted from related works using the same data split [2].
Linear, Sparse, and Dense (R2): These represent different sampling mask schedules for degradation. Experiments show that the Dense schedule yields better results, indicating the early recovery phase critically influences final reconstruction quality.
Algorithm 2 (R3): Since k_T is a subset of k_c, we define k_T=k_c*M_T.
Multi-coil Data (R3): Our method can be extended to multi-coil scenarios by increasing the input channels and incorporating coil sensitivity maps as additional conditions.
Theoretical Justification (R3): Our method can be viewed as a special case of diffusion bridges. As shown in FDB[1] (Eq. 10), setting w_t=0 recovers our approach. Other Questions (R2): All minor writing issues will be corrected in the final version. [1] Learning fourier-constrained diffusion bridges for MRI reconstruction. [2] Cycle-Consistent Bridge Diffusion Model for Accelerated MRI Reconstruction. [3] CDiffMR: Can we replace the gaussian noise with k-space undersampling for fast mri?
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
This paper proposes FilterDiff, a deterministic k-space diffusion framework for accelerated MRI reconstruction. The approach differs from conventional noise-based diffusion models by adopting a noise-free, undersampling-based degradation and reconstruction trajectory. Empirical results on two benchmark datasets show promising improvements over several baselines. The authors also introduce custom k-space loss functions and a reconstruction process tailored to the proposed framework.
During the initial review phase, concerns were raised by two reviewers (R1 and R2) about the paper’s novelty and its insufficient engagement with closely related prior work, particularly CDiffMR (MICCAI 2023) and FDB. Specifically, reviewers noted that similar ideas — k-space degradation, deterministic diffusion trajectories, and bridge-based modeling — had already been presented in these works. Reviewer 1 also emphasized that claims about novelty and distinctiveness from FDB were inaccurate. R2 highlighted the lack of citation and direct experimental comparison to CDiffMR.
In the rebuttal, the authors clarified that FilterDiff differs by starting the diffusion process from a low-degraded state (rather than a severely degraded one) and performing degradation and recovery entirely in the k-space domain with a custom loss and reconstruction function. They also acknowledged the oversight in not citing CDiffMR and promised to correct this in the final version. However, the rebuttal reiterates novelty claims without fully acknowledging the conceptual overlap with FDB, which already uses noise-free, k-space-based diffusion bridges with deterministic degradation. Reviewer 1 maintained their rejection after rebuttal due to these unresolved concerns, viewing the novelty as insufficient for publication.
Reviewer 3 viewed the deterministic design and empirical performance as promising, ultimately recommending acceptance. However, their concerns primarily centered around implementation clarity rather than conceptual originality.
Given the centrality of novelty and proper contextualization in methods contributions at MICCAI, and the fact that the rebuttal did not convincingly dispel key doubts raised by expert reviewers, I recommend rejection of this submission. The authors may consider more thoroughly revisiting the technical distinctions with FDB and CDiffMR, potentially reframing the work around empirical or architectural contributions rather than methodological novelty.
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A