List of Papers Browse by Subject Areas Author List
Abstract
Multi-parametric magnetic resonance imaging (MRI) is an advanced MRI technique that can provide multiple quantitative maps simultaneously based on acquired multi-echo images. However, the lengthy scan time often limits its application. Accelerated multi-parametric MRI using deep learning is of great interest. The existing studies have two limitations: 1) inefficient use of the multi-echo information; 2) lack of physical prior for parametric mapping. To address these issues, in this work, we propose a novel decoupling-driven and physics-informed reconstruction network for accelerated multi-parametric MRI. Specifically, to better align and integrate multi-echo information, we propose a novel decoupling technique consisting of wavelet-driven decoupling module, contrastive and echo-dependent decoupling losses, such that the multi-echo features can be effectively decoupled into echo-dependent and echo-independent components. Only the echo-independent features are fused across multiple echoes. Besides, Bloch equations are incorporated as physical priors to guide the parametric mapping network. Experimental results on our in-house data (12-echo sequence) show that our method outperforms the state-of-the-art methods by 1.54% in average SSIM and 1.70dB in average PSNR for 4× acceleration, which significantly advances the performance limitation for multi-parametric MRI. Our code is available at https://github.com/IDEARL23/WDPM-Net.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3953_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/IDEARL23/WDPM-Net
Link to the Dataset(s)
N/A
BibTex
@InProceedings{DanRui_Waveletdriven_MICCAI2025,
author = { Dan, Ruilong and Sun, Kaicong and Zhou, Yichen and Jia, Minqiang and Liu, Yuxuan and Zhang, Han and Zong, Xiaopeng and Shen, Dinggang},
title = { { Wavelet-driven Decoupling and Physics-informed Mapping Network for Accelerated Multi-parametric MR Imaging } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15960},
month = {September},
page = {666 -- 675}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper seeks to generate quantitative parameter maps from a multi-echo MRI sequence using a new end-to-end neural network architecture. The pulse sequence is a multiecho gradient echo (MULTIPLEX) that encodes T2* and T2. The proposed architecture starts with undersampled k-space data and generates the final parametric maps, but also reconstructs the multiecho images as an intermediate step. These images are then processed with conventional analytical fitting to estimate parameter maps, which are used as input into the final parameter estimation network in a physics-informed manner. The results are compared quantitatively and qualitatively with in vivo data to existing methods and found to perform better.
The main contributions of this paper include the overall end-to-end training strategy, a sophisticated reconstruction module using decoupling and fusion of multi-echo image information, and a physics-informed network for parameter estimation. While the network is intended to be trained end-to-end, the two components (reconstruction and parameter mapping) can be used independently and combined with other methods.
- 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 overall architecture is innovative. The reconstruction component takes advantage of the multi-echo structure. Generating the intermediate results (multiecho images) and including them in the loss function should help guide the solution to physically reasonable results. 2) The network performs well, based both on the quantitative and the invivo examples shown in Fig 2.
- 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.
Weaknesses of this paper are 1) the physical priors for T1 and T2 that are introduced are both linear approximations of nonlinear processes, which may not be clear to readers unfamiliar with relaxometry, and may introduce bias especially in low SNR data. Additionally, 2) the generalizability of this approach is unclear. Presumably both networks would need to be retrained on new data if applied to data acquired at other sites or with different parameters. Finally, 3) the experimental undersampling is simulated retrospectively rather than a prospective demonstration of actual acceleration.
- 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
While this network is trained end-to-end, the two steps can be treated and trained fully independently. The ablation studies partially demonstrate this, and future studies could more completely evaluate the separate components. While there are innovations in both stages (recon and mapping) it is not clear which ones are most valuable, and it would be easier for others to understand the full value of these innovations if they were evaluated in separate papers. For example, there are many DL reconstruction methods that could be used in place of step 1. JUST-Net is an appropriate example, and including it is a strength of this work, but they are other alternatives. The proposed wavelet-driven decoupling recon could be compared directly with these independent of the mapping step. Due to the independence of the two steps it would be possible to perform many variations in an ablation study. The small ablation matrix sampled here is valuable (and appropriate in scope for MICCAI) but incomplete for understanding the value of the innovations.
Some specific comments to consider in rebuttal: 1) The paper does not describe the GT used for training. I infer that the GT is the fully sampled images reconstructed by the vendor, but this should be clarified. 2) The conclusion states the improvement over SRM-NET is a large margin, which seems a little overstated. 3) Were sensitivity maps and/or calibration scans used in the reconstruction? It seems not, that the reconstruction was fully data-driven and didn’t spatial encoding information.
- 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 architectural innovations appear novel, but it is difficult to evaluate the relative benefit of each advancement.
- 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.
My concerns from the original review largely remain after rebuttal, focusing on generalizability and uncertainty of the additive value for the different components of the method.
From the rebuttal, I don’t understand the response regarding GT and coil profiles. If the vendor’s images were used as GT, it is not clear what the ESPRIT-derived sensitivity profiles were used for. The overstated advantage over SRM-NET remains.
Review #2
- Please describe the contribution of the paper
The authors present WDPM-Net for accelerated multi-parametric MRI, with the following innovations: 1) Decoupling the multi-echo information into echo-independent and echo-dependent features via wavelet transform and custom-designed losses. 2) physic-informed neural network in which Bloch-equation is employed to produce initial estimates of T1 and T2* maps. WDPM-Net consists of two sequential network modules for image reconstruction and parameter mapping, respectively, and is trained in an end-to-end manner. Evaluation on datasets acquired using MULTIPLEX sequence suggest the method’s outperformance over existing neural networks.
- 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 separation of echo-dependent and echo-independent features represents a novel approach.
Combining echo-independent features across multiple echoes enhances the network’s performance.
The wavelet-driven decoupling enables the formulation of an objective function that incorporates both echo-dependent loss and contrastive decoupling loss, contributing to improved performance.
The use of analytically derived parametric maps as initial estimates appears to guide the network parameters in the right direction.
- 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 proposed method was evaluated solely on datasets acquired using the MULTIPLEX sequence, which is a relatively recent advancement in the field. This raises concerns about the network’s generalizability to more conventional sequences, such as inversion recovery (IR)-based T1 mapping or multi-spin-echo (MSE)-based T2 mapping. The authors are encouraged to discuss or provide experiments on standard sequences to establish broader applicability.
The validity of the separation between echo-dependent and echo-independent features in the WD module is not fully established. It would be valuable to include quantitative or qualitative metrics to verify this separation. Additionally, it is unclear whether the wavelet transform is indeed the optimal choice for this decoupling task. A comparison with alternative decomposition techniques could strengthen this claim.
The normalized root-mean-square error (NRMSE) should be included as an evaluation metric, as it is widely used and offers a more interpretable measure of reconstruction quality.
The impact of the joint end-to-end training of the reconstruction and mapping networks should be further examined. Specifically, it would be insightful to assess how the performance changes when the two networks are trained independently. This analysis could help clarify the benefits of the proposed joint training approach.
The order of references in the manuscript appears to be inconsistent. The authors should revise the reference list to ensure correct and sequential citation formatting.
- 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.
(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?
This paper presents a novel approach through the concept of echo-decoupling and its integration into the design of loss functions. The method shows promise and, with further validation across a broader range of parametric mapping scenarios, is expected to contribute meaningfully to the field. It is hoped that the authors will address the critiques noted above during the rebuttal process.
- 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 and clarified most of my concerns. Thank you.
Review #3
- Please describe the contribution of the paper
The authors proposed a multi-echo image feature disentanglement and physics-based parameter mapping framework to enhance k-space reconstruction and downstream parameter mapping accuracy. The multi-echo feature disentanglement approach shows potential for extension to other qMRI sequences involving multi-echo acquisitions.
- 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 authors proposed a framework to disentangle multi-echo image features in quantitative MRI (qMRI) using the MULTIPLEX sequence, which may be generalizable to other qMRI pulse sequences involving multi-echo acquisitions.
- Both the ablation study and comparative experiments are comprehensive, demonstrating the superior performance of the proposed method.
- 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.
- No visualizations of the decoupled features or attention maps were presented, making it difficult to understand the behavior of the proposed decoupling process, particularly in Equation 1.
- The authors stated that they used equispaced sampling masks; however, the performance under variable density sampling patterns remains unknown, which weakens the rigor of the study.
- The proposed “Physics-informed Mapping Network” appears to be a relatively straightforward implementation—essentially concatenating the analytical solution with a U-Net for parameter estimation—making it less novel than the feature decoupling component.
- As noted previously, the improvement from using the physics-informed mapping is minimal, as shown in the ablation results in Table 2, further reducing the added value of this component.
- The method incorporates multiple loss functions (Eqs. 7–10) with corresponding weighting hyperparameters, but these were only empirically set. A more rigorous hyperparameter tuning procedure is needed to support the robustness of the results.
- 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
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?
The paper proposes a promising framework for multi-echo feature disentanglement and physics-informed parameter mapping in qMRI. The methodology is well-motivated and shows improved performance through comprehensive experiments and ablation studies. However, the lack of visualization for feature decoupling, limited evaluation on sampling patterns, and minimal impact of the physics-informed component slightly reduce the overall novelty and clarity. Despite these limitations, the core idea is valuable and merits acceptance.
- 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 are encouraged by all reviewers’ recognition of our paper’s novelty and clarity. We highly appreciate their valuable comments. We will release the code once accepted. Below are our responses. 1.Model generalizability (R1C1&R3C2): In the experiments, our WDPM-Net achieves promising results for dealing with 12-contrast MTP sequence, an advanced sequence with dual FAs and dual TRs dedicated for multi-parametric mapping. Due to limited pages, we haven’t evaluated the less complex conventional sequences, but our model is supposed to achieve promising performance as well. Generalizability is a common issue in deep learning. Most of the reconstruction methods need to be retrained on different centers, and our model needs to be as well. We will improve our model’s generalizability and applicability in future work. 2.Validity of feature decoupling (R1C2&R2C1): In fact, the decoupling is ensured by our design in Eq. (1), and our echo-dependent and contrastive decoupling losses, which can actually be regarded as quantitative metrics of decoupling. As shown in Table 2, both reconstruction and mapping stages enjoy the efficacy of decoupling, and the mapping’s SSIM increases from 0.8074 to 0.8570. Due to pages, we have not shown the visualization of decoupled features. We will show the t-SNE analysis in future work. Moreover, we found that wavelet-driven decoupling achieves better performance than other separation strategies (e.g., via echo-independent and dependent encoders). Due to pages, we only presented the best strategy we found. 3.Validity of joint training (R1C4&R3 Additional comments): Empirically, we found that joint optimization can lead to better performance for final mapping compared to separate training, although it could trade off the intermediate reconstruction performance. Due to pages, we only show the results of joint training. In the experiment, we have shown that our mapping module PIMN can be easily combined with off-the-shelf reconstruction models (e.g., JUST-Net) for parametric mapping and shows promising mapping performance. We will demonstrate the effectiveness of separate parts more completely in future work. 4.Specific comments from R3: In practice, we got the GT from the vendor software. Our model took the coil-combined data as input, with sensitivity maps calculated by ESPIRIT. We appreciate the reviewer’s comment on statement of comparison with SRM-NET and will clarify all these specific comments in the final version. 5.Undersampling mask and prospective evaluation (R2C2&R3C3): Due to pages, we only investigated the widely used equispaced mask. According to our empirical experience, it should work well for variable density mask as well. We appreciate the advice of prospective study and we have completed it now. Due to pages, we will put it in future work. 6.Effectiveness of the PIMN and physical priors (R2C3&R2C4&R3C1): As discussed in Sec 2.3, Bloch equation-based mapping is sensitive to data quality. We used it as network input and it will be refined by the UNet to add nonlinearity and improve robustness to low SNR scenarios. Table 2 shows that our PIMN improves the SSIM from 0.8570 to 0.8745, indicating its efficacy for the mapping stage. Unlike previous methods which incorporate Bloch equations in loss functions, we are the first to combine Bloch equation-based analytical solutions with data-driven model to effectively integrate linear prior into nonlinearity network. In future work, we can extend our framework by replacing the UNet with more advanced unrolling-based models. 7.Metrics and order of references (R1C3&C5): Thanks for the comments. Since the calculation of NRMSE is related to PSNR, we haven’t included it due to pages. We will compare it in future work. And we will fix the references in the final version. 8.Hyperparameters (R2C5): In the experiments, the optimal hyperparameters were determined by grid search within a standard training paradigm. Due to pages, we have not shown them.
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’
The authors adequately addressed the reviewers’ concerns. There were some lingering concerns about the ground truth, but this seems to be a misinterpretation that was explained in rebuttal.
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’
Most major concerns have been addressed; however, some questions remain and further clarification is needed.