Abstract

Magnetocardiography (MCG) offers superior spatial resolution and sensitivity over electrocardiography (ECG) for non-invasive cardiac assessment, yet clinical adoption is limited by costly, complex superconducting quantum interference devices (SQUIDs) and optically pumped magnetometers (OPMs). Tunneling magnetoresistance (TMR) sensors provide a room-temperature, cost-effective alternative, but their elevated intrinsic noise impedes clinical translation. We propose DFuse, a dual-branch adaptive fusion network for multichannel MCG denoising that jointly captures long-range temporal dependencies and preserves diagnostically critical morphological features under cross-channel heterogeneous signal-to-noise ratio (SNR). Evaluated on a self-developed 24-channel TMR system, DFuse achieves SNR improvements of 20.31 dB on simulated data (4,800 ten-second samples, 12 channels) and 15.79 dB on recordings from 22 healthy volunteers. The residual denoising error, converted to an equivalent noise level of $6 \pm 6\,\mathrm{fT}/\sqrt{\mathrm{Hz}}$ at 1 Hz, approaches the sensitivity range reported for clinical-grade SQUID and OPM systems. These results demonstrate that deep learning–enhanced TMR-based MCG enables accessible, high-fidelity cardiac diagnostics in resource-constrained settings. Code is publicly available at https://github.com/HHarr1son/DFuse.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0421_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/HHarr1son/DFuse

Link to the Dataset(s)

https://huggingface.co/datasets/lhaohaoer/Muti_mcg_dataset

BibTex

@InProceedings{HuaShu_DFuse_MICCAI2026,
        author = { Huang, Shuai AND Li, Hao AND Li, You AND Zhou, Jingbo AND Cui, Jian AND Zheng, Zhong AND Zhang, Xin AND Jiang, Tianzi},
        title = { { DFuse: A Dual-Branch Adaptive Fusion Network for Multi-Channel MCG Denoising } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16896},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper addresses the high noise level in multi-channel magnetocardiography (MCG) signals acquired by TMR sensors, which limits their clinical applicability. The authors propose DFuse, a dual-branch adaptive fusion network that decouples long-range temporal context modeling and morphology-aware representation, and integrates them through time-varying and sample-wise adaptive fusion. Additionally, the paper constructs both simulated and real multi-channel MCG datasets and demonstrates superior performance under a leave-one-subject-out protocol. The method significantly improves SNR and reduces the noise floor to a level comparable to clinical-grade devices, highlighting the potential of low-cost TMR-based MCG systems for practical applications

  • 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 experimental setup is comprehensive, including simulated and real datasets, comparisons with multiple baselines, and ablation studies. 2.This paper is well-written and easy to follow, with a clear structure and smooth logical flow.

  • 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 model is validated only on healthy subjects; its robustness on pathological signals remains unclear. 2.The interpretability of the model, especially the fusion mechanism, is not sufficiently analyzed. 3.The table formatting should be further standardized and aligned with the template. In addition, in Table 1, the term “Stimulate” (referring to simulated data) is incorrectly used and should be corrected to “Simulated”. It is recommended to ensure consistent terminology throughout the manuscript.

  • 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.

    (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?

    This paper addresses a practically important problem of denoising multi-channel MCG signals acquired by TMR sensors. It proposes a well-motivated and reasonably designed dual-branch adaptive fusion network. The experimental evaluation is comprehensive, demonstrating consistent performance improvements on both simulated and real datasets. Notably, the method reduces the noise level of low-cost TMR systems to a range comparable to clinical-grade devices, indicating promising practical potential.However, several limitations remain. The model is validated only on healthy subjects, and its robustness on pathological signals remains unclear. In addition, the interpretability of the model—particularly the adaptive fusion mechanism—is not sufficiently analyzed. Furthermore, the presentation quality could be improved: the table formatting should be better aligned with the template, and in Table 1, the term “Stimulate” is incorrectly used instead of “Simulated”; consistent terminology should also be maintained throughout the manuscript.

  • 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.

    After reading the rebuttal, I maintain my original score. The paper addresses an important and practically relevant problem: improving the signal quality of low-cost TMR-based multi-channel MCG. The proposed dual-branch adaptive fusion framework is reasonably motivated, as it separates long-range contextual modeling from morphology-aware feature extraction and further exploits cross-channel information. The experiments on both simulated and real TMR-MCG data also provide useful evidence for the effectiveness of the method. The authors have addressed my main concerns to a reasonable extent. The lack of pathological validation remains a limitation, but it is clearly acknowledged. The explanation of the time-varying fusion gate partially improves the interpretability of the fusion mechanism, and the proposed heatmap visualization would further help clarify this point. However, I still think the presentation quality needs careful improvement. The rebuttal itself contains LaTeX/Markdown-style notation such as “$A_\ell$” in a plain-text response box, where it is not rendered properly and reduces readability. This further suggests that the authors should pay more attention to formatting and proofreading in the final version. Overall, although some limitations remain, I believe the paper’s contribution outweighs these issues. I therefore maintain my original score.



Review #2

  • Please describe the contribution of the paper

    This paper tackles an important problem in TMR-based multi-channel MCG, where the low cost and room-temperature operation of TMR sensors come with the downside of substantial intrinsic noise. To address this, the authors propose DFuse, a dual-branch denoising framework that separately models long-range contextual information and fine waveform morphology, and then fuses the two adaptively. The paper also introduces both a simulated benchmark and a real 24-channel TMR-MCG dataset, and shows strong denoising performance under LOSO evaluation.

  • 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 addresses the problem itself is genuinely meaningful. MCG is clinically promising, but systems based on SQUID or OPM are still expensive and cumbersome to deploy. From that perspective, improving the signal quality of low-cost TMR-based MCG is not just a technical exercise, but something that could actually matter in practice.

    The proposed method is also quite intuitive. It makes sense to separate long-range rhythm/context information from local morphology-related information, rather than expecting a single branch to handle both equally well.

    Also, I think the experimental section is fairly solid overall. The paper evaluates the method on both simulated and real data, compares against a good number of baselines, and includes ablation studies that help support the design choices.

  • 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.

    I have several concerns regarding the evaluation. Most importantly, the real-data reference is derived via cycle-averaging (101 cycles) from the noisy signal itself. This is problematic because it’s unclear whether DFuse is truly denoising the signal or simply learning to approximate a smoothed template. Without an independent clean reference, the performance gains on real data are difficult to validate.

    I also had some reservations about the simulated benchmark. The paper rightly points out that ECG and TMR-based MCG have different signal and noise characteristics, but the simulated data are still derived from PTB-XL ECG signals injected into a coil-based setup. This may be a practical approximation, but I am not fully convinced that it captures the complexity of real multi-channel MCG.

    Another point is that MGU-Net is discussed in the introduction as a relevant prior method for TMR-based MCG denoising, yet it is not included in the comparisons. I understand that it is a single-channel method, so the comparison would not be perfectly matched, but some form of discussion or adapted baseline would still have been useful.

    I also felt that the method section could describe the architecture in a more concrete way, since some parts remain a bit high-level for reproducibility.

    Finally, the paper is generally readable, although some figures are dense and the font size is too small in a few places.

  • 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.

    (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?

    I ended up leaning positive on this paper because the problem is meaningful and the overall method is intuitive. Improving denoising for low-cost TMR-based MCG feels practically relevant. My main hesitation came from the real-data evaluation setup and from some uncertainty about how realistic the ECG-derived simulated benchmark is. So overall, I saw this as a good and useful paper.

  • 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.

    I appreciate the authors’ response. Some clarifications were helpful, but my overall opinion remains unchanged. I maintain my initial recommendation.



Review #3

  • Please describe the contribution of the paper

    The main contribution of this paper lies in the development of DFuse, a novel dual-branch adaptive fusion network for multi-channel TMR-MCG denoising, which decouples long-range contextual modeling from morphology-sensitive feature extraction and further integrates them with time-varying/sample-adaptive fusion and cross-channel reweighting.

  • 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 so-called innovation of this work largely remains at the level of engineering-style integration of existing dual-branch encoding, attention-based fusion, and channel reweighting modules, with only task-specific adaptation for TMR-MCG denoising.

  • 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 overall writing is not sufficiently concise, which makes the presentation of the method motivation, the relationships among modules, and the core innovations insufficiently clear, thereby affecting readability. In addition, the manuscript still has some issues in terms of English writing quality and consistency, such as inconsistent capitalization, hyphenation, and terminology usage. 2) From a methodological perspective, the architectural contribution of this work appears rather limited. It is closer to a combinational integration of common components, such as dual-branch encoding, attention-based fusion, and channel reweighting, rather than a genuinely novel core mechanism. Although the authors apply the framework to the TMR-MCG denoising scenario, such task-specific adaptation alone is not sufficient to support a strong methodological contribution. This is especially the case given that TMR-based magnetocardiography has already seen continued research progress in recent years, while denoising itself is no longer a particularly frontier research topic. 3) The experimental comparison is still insufficient. The current baselines mainly include several deep learning models and conventional FIR/IIR and DWT methods, but coverage of analytical denoising approaches commonly used in clinical or engineering practice in the MCG/MEG field is clearly inadequate. At a minimum, the authors should discuss or compare against projection- and subspace-based methods (e.g., SSP, SSS, HFC, and tSSS), as well as filtering strategies based on reference sensors or environmental noise modeling. Compared with these methods, which are more commonly used in practice and have clearer physical interpretability, deep learning approaches are often less explainable. Therefore, comparison against only a limited set of baselines is not sufficient to fully demonstrate the superiority of the proposed method. 4) The paper also shows clear weaknesses in reproducibility. The manuscript does not clearly specify the loss function used for training, although this is a critical piece of information for any deep learning-based denoising method, and no explicit formula or description can be found in the method section. At the same time, important implementation details, such as training hyperparameters, optimizer settings, learning rate schedule, batch size, number of training epochs, and model selection criteria, are not sufficiently described, making it difficult for readers to reproduce the reported results. In addition, no code or pretrained model appears to be publicly available, which further weakens the verifiability of the work. 5) In Fig. 1, where is Fltw actually output? I could not identify its location in the diagram, although I later found from the text how it is derived. This point should be clarified more explicitly in the figure or in the accompanying explanation. 6) The core application goal of magnetocardiography is disease screening and pathological feature identification. However, the experiments in this paper are restricted to healthy subjects, and both the simulated and real-data experiments are mainly based on normal waveforms. For clinically more important abnormal or distorted waveforms, such as arrhythmias, conduction abnormalities, or ischemia-related changes, no targeted validation is provided. Therefore, it remains unclear whether the proposed denoising method may distort, attenuate, or overly smooth diagnostically meaningful pathological features. This limitation directly affects the credibility of the method for clinical application. Although the authors acknowledge in the conclusion that the method has only been validated on healthy subjects and that its ability to preserve pathological features remains unverified, I believe this issue is more than a general limitation; it is a fundamental weakness that significantly undermines the external validity of the paper’s conclusions. 7) The statement in the abstract that “The resultant equivalent noise floor on in vivo data reaches 6±6 fT/√Hz at 1 Hz on the real dataset—comparable to clinical-grade SQUID and OPM systems” should be treated with caution. This wording may easily give the impression that the system performance in this work is already close to that of clinical-grade SQUID/OPM systems, but whether such a claim is fully justified requires much more rigorous explanation. Although the manuscript mentions that the residual noise level is obtained “by converting SSD to noise PSD at 1 Hz following [17],” the paper does not provide sufficiently clear details regarding the calculation procedure, underlying assumptions, or physical interpretation. Considering that real magnetic noise is typically dominated by pronounced 1/f characteristics, and that thermal noise, environmental noise, and system noise have a much more complex composition, directly converting denoising error into an equivalent noise level and then comparing it to the intrinsic sensitivity floor of SQUID/OPM systems may be somewhat overstated. The authors should clearly explain the precise definition of this metric, the conversion procedure, the spectral assumptions involved, and its relationship to the true sensor noise floor, in order to avoid potentially misleading statements. [1] Jia Y, Pei H, Liang J, Zhou Y, Yang Y, Cui Y, Xiang M. Preprocessing and Denoising Techniques for Electrocardiography and Magnetocardiography: A Review. Bioengineering (Basel). 2024 Nov 2;11(11):1109.doi: 10.3390/bioengineering11111109.PMID: 39593769; PMCID: PMC11591354.

  • Please rate the clarity and organization of this paper

    Poor

  • 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.

    (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?

    I gave this paper a relatively low score because its methodological novelty appears limited, with the proposed framework being largely an engineering-style combination of existing dual-branch, fusion, and reweighting modules rather than a clearly new core mechanism. In addition, the paper has important weaknesses in reproducibility and validation, including missing loss-function/training details, limited baselines, and evaluation only on healthy subjects without evidence that clinically meaningful pathological MCG features can be preserved after denoising.

  • 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 the meta-reviewer and all reviewers. We address all concerns below and will incorporate clarifications into the revision.

Reviewer 1 Q1:Lack of pathological data Due to the lack of publicly available pathological TMR-MCG datasets, experiments are conducted on healthy subjects only. Evaluation on patient cohorts is planned as future work. Q2:Fusion interpretability The time-varying gate $A_\ell$ produces per-timestep branch weights where $A_{\ell,2}$ is consistently higher around QRS complexes and lower during baseline segments. We will add a heatmap visualization in the revision. Q3:Presentation All formatting issues will be corrected in the final version.

Reviewer 2 Q1:Cycle-averaging as ground truth As described in [7], cycle-averaging is standard practice in TMR-MCG signal acquisition. If DFuse learned a smoothed template, it would not outperform all baselines trained with the identical reference. The gap over the second-best method demonstrates genuine signal recovery. Q2:Simulation realism PTB-XL waveforms are injected into a Helmholtz coil platform replicating the TMR sensor array and analog front-end, with noise directly acquired by TMR hardware. The simulated dataset therefore contains true signal labels (original waveforms before noise injection) and authentic TMR noise, validating two capabilities: whether DFuse learns global structured patterns rather than fitting arbitrary noise, and whether it suppresses TMR-specific noise while preserving morphology. Q3:Missing baselines We have conducted additional comparisons with SSP and MGU-Net and will include these results in the revision. Q4:Reproducibility Key details: MSE loss, AdamW, batch size 16, Caputo orders ${0.3,0.5,0.7,1.0}$. Full implementation details, code, and datasets will be released upon acceptance. Q5 follows R1 Q3. Reviewer 3 Q1:Writing Writing and terminology issues will be corrected throughout the revision. Q2:Novelty and motivation The unifying motivation of DFuse is spatially heterogeneous SNR across TMR-MCG channels: sensors close to the cardiac source capture high-amplitude signals while distant sensors are noise-dominated. A single encoder cannot simultaneously serve both regimes, inevitably trading off between structural coherence and local morphological detail. This directly motivates the dual-branch design: the Context-Aware Encoder provides a global rhythm scaffold robust to SNR variation, while the Morphology-Aware Encoder preserves fine-grained waveform transitions most at risk in low-SNR channels. Caputo fractional derivatives are chosen here because they are the only differentiable operator family continuously interpolating between trend integration and sharp transition detection via $\alpha$; Having two branches then naturally raises the question of how to combine them: different channels and cardiac phases require different feature mixes, motivating the time-varying and sample-wise gates in the Enhanced Fusion Block. CCR in the decoder closes the loop by leveraging reliable high-SNR channels to guide reconstruction of weaker ones. The ablation in Table2 confirms each component is necessary. Q3:Missing baselines The dominant noise in TMR-MCG is intrinsic 1/f and thermal sensor noise rather than environmental interference. SSS and tSSS assume spatial array completeness incompatible with our 5$\times$5 grid. We have conducted additional comparisons with SSP and MGU-Net and will report these results in the revision. Q4 and Q5 follow R2 Q4.Q6 follows R1 Q1.Q7:Noise floor claim Following [17], SSD is first converted to RMS noise in the time domain, then mapped to PSD by accounting for effective bandwidth, enabling direct comparison with sensor noise specifications in pT/$\sqrt{\text{Hz}}$. This estimates residual denoising error in field units rather than intrinsic sensor sensitivity. A derivation will be included and abstract wording revised accordingly.




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 addresses an important problem of denoising multi-channel MCG signals acquired by TMR sensors. it is generally easy to follow. However, the design choices for each module need clearer justification. The descriptions of the context encoder and morphology-aware encoder lack sufficient detail for reproducibility, and the fusion module is only briefly described, which may hinder understanding. The authors should also clarify how the ground truth is obtained. Please refer to the reviewers’ comments for additional suggestions.

  • 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 paper received mixed scores. After the rebuttal, two experts are happy with the clarifications, while one expert still has some concerns regarding the potential clinical impact. It can be accepted due to the new application rather than the novelty in terms of theoretical contribution.



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.

    The paper addresses the practically relevant challenge of denoising multi-channel MCG signals from low-cost TMR sensors using a dual-branch adaptive fusion network. While the authors’ rebuttal clarified important implementation details and added requested baseline comparisons, reviewers remained concerned about the incremental methodological novelty and the critical lack of validation on pathological data. Because the method was only evaluated on healthy subjects—leaving it unclear whether diagnostically important abnormal waveforms (e.g., arrhythmias) might be distorted or smoothed out—the paper currently lacks the robust clinical validation necessary for acceptance.



Meta-review #3

  • 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.

    The rebuttal provided useful clarifications; however, several key concerns raised by the reviewers were not fully resolved. As a result, the paper does not reach the acceptance threshold for MICCAI.



back to top