List of Papers Browse by Subject Areas Author List
Abstract
Magnetic Particle Imaging (MPI), an emerging technique with high sensitivity and resolution, requires time-consuming calibration for System Matrix (SM)-based reconstruction. Due to the strong locality and redundancy in the frequency domain, sparse sampling can capture sufficient information for rapid SM calibration without full-size SMs. However, it often leads to low-frequency energy leakage due to nonlinear magnetization of nanoparticles, causing the loss of low-frequency components. These components are essential for maintaining the SM’s shape, and their absence leads to structural degradation and visible artifacts. Current methods tend to overemphasize high-frequency features, neglecting these low-frequency ones. Besides, single-step upsampling leads to error accumulation, especially with large scaling ratios, degrading reconstruction quality. To address these issues, we propose the Iterative Frequency Restoration-Fusion Network (IFRFNet), which uses an iterative frequency-domain restoration-fusion module. Unlike single-step upsampling, our approach refines, fuses, and upsamples high- and low-frequency features in stages, ensuring continuous optimization. This prevents error accumulation, preserves fine details, and maintains structural integrity. By iteratively recovering low-frequency components and refining high-frequency details, IFRFNet minimizes artifacts and retains crucial information. The Effective Upsampler further enhances the quality of the features, ensuring clear and realistic final SM volumes. Experiments on the OpenMPI dataset show that IFRFNet achieves SOTA performance.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2466_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{XuWei_IFRFNet_MICCAI2025,
author = { Xu, Weixin and Zhai, Penghua and Tian, Jie and Mu, Wei},
title = { { IFRFNet: Iterative Frequency Restoration-Fusion Network for Fast System Matrix Calibration on Magnetic Particle Image } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15962},
month = {September},
page = {295 -- 304}
}
Reviews
Review #1
- Please describe the contribution of the paper
The main contribution of the paper is the development of the Iterative Frequency Restoration-Fusion Network (IFRFNet) for improving System Matrix (SM) calibration in Magnetic Particle Imaging (MPI). Unlike conventional methods that overemphasize high-frequency features and suffer from low-frequency energy leakage due to nonlinear magnetization, IFRFNet iteratively restores and fuses both high- and low-frequency components in the frequency domain. This staged refinement approach prevents error accumulation, preserves structural integrity, and reduces reconstruction artifacts.
- 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.
One major strength of the paper is the novel formulation of IFRFNet, which effectively addresses the common issue of low-frequency energy leakage in System Matrix (SM) calibration for Magnetic Particle Imaging. Unlike existing methods that primarily focus on high-frequency features, IFRFNet introduces a stage-wise, iterative restoration and fusion strategy in the frequency domain that recovers lost low-frequency components while refining high-frequency details. This approach leads to slight improvements over previous methods in terms of reconstruction quality.
- 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.
One major weakness of the paper is that the performance improvement over previous methods appears to be relatively small. While the proposed approach is conceptually interesting, it is unclear whether the gains are substantial enough to provide a practical advantage in real-world MPI applications. Secondly, the paper lacks an ablation study to support its core claims. Without such an analysis, it is difficult to determine which components of the proposed IFRFNet architecture contribute to the observed improvements. In particular, the claim that low-frequency components are better preserved would benefit from more direct and quantitative evidence. For example, frequency-domain comparisons or visualizations could help demonstrate the effectiveness of the proposed restoration-fusion process. Lastly, the introduction does not include any citations when discussing MPI and MPI reconstruction, which weakens the context and motivation of the study.
- 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.
- 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?
While the proposed method presents a conceptually interesting approach to frequency restoration in MPI, the actual performance improvements over prior work are relatively limited. The absence of ablation studies and quantitative evidence for key claims, particularly regarding low-frequency recovery, makes it difficult to assess the effectiveness of the method.
- 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 adequately addressed the concerns raised during the review through their rebuttal. I believe the revisions sufficiently resolve the issues, and I recommend acceptance.
Review #2
- Please describe the contribution of the paper
This paper proposes an Iterative Frequency Restoration-Fusion Network (IFRFNet) for reconstructing high-resolution system matrices (SMs) from their low-resolution counterparts, addressing the time-consuming calibration process inherent in high-resolution MPI scanning. Unlike previous state-of-the-art (SOTA) methods, the key motivation behind IFRFNet is to preserve critical low-frequency information through a combination of a Frequency Attention Module, a Frequency Filter and Refine-Fusion Module, and an iterative frequency restoration architecture.
- 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 research direction is highly relevant and significant for real-world clinical applications.
-
The proposed IFRFNet demonstrates superior performance compared to previous state-of-the-art 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.
-
The authors claim that previous methods tend to lose low-frequency features; however, this point is not clearly explained or sufficiently justified in the paper.
-
While the authors state that their method preserves both low- and high-frequency information, the underlying mechanisms and evidence supporting this claim are not clearly presented.
-
The paper introduces a Frequency Filter module based on CNN layers to extract low- and high-frequency features. However, it remains unclear why the extracted features can be definitively categorized as low- or high-frequency. Could the authors provide further insights or theoretical justification for this distinction? For example, why can these features not be swapped? In addition, visualizations would be helpful to support this claim.
-
The authors argue that “single-step upsampling methods suffer from error accumulation” and that their iterative structure alleviates this issue. However, from my perspective, iterative methods may instead exacerbate error propagation unless carefully designed. More clarification and empirical justification are needed.
-
No ablation studies are provided to analyze the effectiveness and individual contributions of the proposed modules. Including such studies would strengthen the validity of the proposed 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 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.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
As metnioned above.
- Reviewer confidence
Somewhat confident (2)
- [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 reviewing the other reviewers’ comments and the authors’ rebuttal, I find that only the low- and high-frequency differences have been adequately addressed, while the remaining concerns remain unresolved.
Review #3
- Please describe the contribution of the paper
This work proposes a novel SM calibration framework for MPI. The frequency domain feature processing module fusion the features with different frequencies to help better restore the information gap between low-resolution and high-resolution SMs.
- 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 presents the proposed method with clear language and logical organization, making the methodology easy to follow and understand.
- The proposed method effectively addresses the limitation of existing super-resolution networks, where the upsampling modules fail to fully optimize the extracted features.
- 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.
lacks visual result figures and comparisons
- 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
- Although space is limited, as a method targeting image super-resolution, it is still important to include qualitative visual results, even if presented in a reduced size. Such examples are essential to demonstrate the perceptual quality and practical effectiveness of the proposed approach.
- The current presentation lacks detailed mathematical expressions for individual modules. A more thorough formulation could be considered in an extended version to improve the rigor and completeness of the method description.
- 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?
The paper clearly outlines the background of MPI calibration and explicitly identifies a key limitation of existing methods—namely, the insufficient capture and utilization of low-frequency components. In response, it proposes an effective solution to address this issue. The manuscript reflects the authors’ deep understanding of the field, strong scientific writing skills, and high level of research competence.
- 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 some of my concerns.
Author Feedback
We appreciate all reviewers’ valuable comments. Here are our responses: Q1 for R2&3: Lacks visual result figures and comparisons. We thank the reviewers for the constructive comments. We performed visual comparison of both system matrix (SM) and reconstruction results, showing our model provides images with higher quality, better structural integrity, and accurate shape. These contents were not included in the main text due to space restriction, we had to prioritize more important content. Q2 for R3: Explanation of the low-frequency features loss was not sufficient. A key characteristic of MPI is its nonlinear magnetization response, which enables imaging but concentrates signal energy in high frequencies, leading low-frequency components relatively weak and easily suppressed[1]. Also, sparse sampling cannot fully cover all spatial points[2]. Both factors lead to low-frequency loss and distort the overall shape of reconstructed images. [1]Magnetic particle imaging: Current developments and future directions [2]Sparse Reconstruction of the Magnetic Particle Imaging System Matrix Q3&Q4 for R3: Mechanism for preserving and distinguishing low- and high-frequency features is unclear. Thanks for the comments. The principle of frequency decomposition was adopted, in which low-frequency features capture the global structure, while high-frequency ones encode local details. Specifically, average pooling is applied to the input feature map FIn to extract the low-frequency component FLow, as the average pooling acts as a low-pass filter that suppresses high-frequency signals while preserving overall structure. The result is then upsampled to match the original resolution. The high-frequency component FHigh is obtained as the subtraction FIn-FLow, which contains the local details filtered out by pooling. Q5 for R3: Improper iterative upsampling may exacerbate error propagation. We agree that improper iterative upsampling may cause error propagation. To avoid any issues, we carefully designed the upsampler. To minimize errors, feature recovery operations are included after the upsampling step. The iterative process gradually refines and fuses shape and detail information by combining Frequency Filter (FF) and Frequency Refine-Fusion Module (FRFM), also preventing error accumulation. Experimental results show that our method achieves SOTA performance, especially at high scale ratios, proving the effectiveness of the design. Q6 for R3&4: No ablation study. We agree that ablation studies help clarify each module’s contribution. IFRFNet is a tightly coupled framework with interdependent modules. For example, FF separates frequency features, which FRFM refines and fuses. Removing FRFM disrupts FF, invalidating the design. Our method relies on iterative frequency separation, optimization, fusion, and upsampling through synergistic modules. Without these, IFRFNet degrades into SRCNN. Experimental results prove the effectiveness of our design (Table 1-3). Q7 for R4: Performance improvement is relatively small. Thanks for the comment. IFRFNet not only designed for improving SM calibration metrics but also to overcome issues in prior work, including neglect of low-frequency information and errors accumulation from single-step upsampling. SM calibration improvement in nRMSE may appear small due to output-range normalization (nRMSE=RMSE/(ymax−ymin)), it remains meaningful. Also, its advantages are more evident in downstream image reconstruction, the ultimate goal in MPI, where it improves PSNR by 1.5 dB and SSIM by 8%, indicating more informative and accurate SMs. Moreover, IFRFNet achieves this with only half the parameters of the second-best model (9.6M vs. 20M), reflecting its efficiency. Q8 for R4: Lacks citations related to MPI and its reconstruction. Thanks for the constructive comment. The literature relevant to MPI and its reconstruction will be added in the final version to enhance background explanation and motivation analysis of the article.
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.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
This paper introduces IFRFNet, a frequency-domain neural network framework for improving system matrix (SM) calibration in Magnetic Particle Imaging (MPI). The method is motivated by the need to address low-frequency information loss and reconstruction artifacts present in previous approaches. IFRFNet introduces an iterative architecture incorporating a Frequency Filter module and a Refine-Fusion strategy designed to restore and preserve both low- and high-frequency features, aiming to mitigate issues such as error accumulation from single-step upsampling. The paper demonstrates promising results in SM calibration and downstream image reconstruction performance, supported by quantitative evaluation across multiple metrics.
The reviewers largely agree on the novelty and clinical relevance of the proposed direction, as well as on the clarity and quality of writing. The methodology is thoughtfully constructed and exhibits modest but consistent improvements over prior work, along with notable parameter efficiency. Reviewer 2 offered a clear accept recommendation, praising the scientific rigor and practical motivation. Reviewers 3 and 4 expressed more cautious support, initially flagging the lack of visual examples, ablation studies, and limited theoretical or empirical justification for some of the design choices—particularly in relation to the frequency filtering mechanism and the iterative refinement strategy. However, both shifted to an accept stance after the rebuttal, which provided reasonable clarifications and additional discussion on design rationale, albeit without new experiments.
While some concerns regarding empirical validation remain—most notably the absence of ablation experiments and visual qualitative results—the rebuttal was comprehensive and addressed major criticisms constructively. Given the paper’s conceptual novelty, relevance to a practical and underexplored area in MPI, and general soundness of the proposed approach, this AC recommend acceptance, with a suggestion that the authors include visual results, citations, and ablation analysis in the final version to reinforce the clarity and reproducibility of the work.
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