List of Papers Browse by Subject Areas Author List
Abstract
Echo Planar Imaging (EPI) is widely used for its rapid acquisition but suffers from severe geometric distortions due to B0 inhomogeneities, particularly along the phase encoding direction. Existing methods follow a two-step process: reconstructing blip-up/down EPI images, then estimating B0, which can introduce error accumulation and reduce correction accuracy. This is especially problematic in high B0 regions, where distortions align along the same axis, making them harder to disentangle. In this work, we propose a novel approach that integrates Implicit Neural Representations (INRs) with a physics-informed correction model to jointly estimate B0 inhomogeneities and reconstruct distortion-free images from rotated-view EPI acquisitions. INRs offer a flexible, continuous representation that inherently captures complex spatial variations without requiring predefined grid-based field maps. By leveraging this property, our method dynamically adapts to subject-specific B0 variations and improves robustness across different imaging conditions. Experimental results on 180 slices of brain images from three subjects demonstrate that our approach outperforms traditional methods in terms of reconstruction quality and field estimation accuracy.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3137_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/wenqihuang/PINR
Link to the Dataset(s)
N/A
BibTex
@InProceedings{HuaWen_PhysicsInformed_MICCAI2025,
author = { Huang, Wenqi and Wang, Nan and Liao, Congyu and Lin, Yimeng and Gao, Mengze and Rueckert, Daniel and Setsompop, Kawin},
title = { { Physics-Informed Implicit Neural Representations for Joint B0 Estimation and Echo Planar Imaging } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15962},
month = {September},
page = {486 -- 496}
}
Reviews
Review #1
- Please describe the contribution of the paper
This work presents a physics-informed neural network for joint estimation of B0 and images. The authors employed rotated-view EPI acquisitions to leverage B0-induced distortions that appear differently across multi-views, and thereby achieving distortion-corrected images as well as accurate B0 field estimation. The authors’ network model maps spatial coordinates to B0 and images, allowing for flexibility in setting rotation angles for EPI acquisitions. The authors evaluated their method’s performance in comparison to conventional TOPUP method with simulated EPI data and also with prospectively acquired EPI 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.
- No need for image interpolation in settings with multi-view EPI acquisitions and distortion correction
- No need for additional calibration scans for B0 estimation
- No increase of scan time despite multi-view EPI scans
- The idea of mapping spatial coordinates to signal values is novel
- The network model does not entail complicated architecture, but seems to learn what it is expected to learn.
- 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.
- Limited evaluation: there have been a number of methods introduced, attempting EPI distortion correction via deep learning. Comparison to real “state-of-the-art” methods is missing.
- Generalizability: The datasets were obtained using EPTI, which is quite sophisticated but not yet an established sequence for EPI imaging. Validation with a generally used EPI sequence would yield more impact to the field.
- Related, if training datasets with a coarser resolution were provided, would the proposed network still be able to capture continuous representation of B0 field?
- The authors stated “a smoothed L1 loss for data consistency”, but according to Eq. (5) it looks like they used L2 loss.
- The number of datasets for testing the method is not given in Table 1.
- What does “m” represent in the symbol for the selection mask? Is it “n”?
- The order of references in the manuscript is mixed.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 presents novelty with the idea of mapping spatial coordinates to signal values and incorporating physical equations in the network. The method, if real ground truth information is provided, is expected to be useful (upon more complete validations). Conversely, since real ground truth is difficult to obtain in this type of imaging scenarios, the method would be rather limited in its use. Nevertheless, in my assessment, the paper’s novelty outweighs its limitations, and thus deserves MICCAI’s attention after 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.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper proposes a Physics-informed Implicit Neural Representation (PINR) to jointly estimate B0 inhomogeneity and reconstruct undersampled images from rotated-view Echo Planar Imaging (EPI) acquisitions, without the need for interpolation due to the use of INR. Experiments using simulated undersampled k-space data show improved performance of the proposed PINR with rotated-view EPI compared to B0 estimation from blip-up/down acquisitions using the conventional TOPUP method and the proposed PINR-based solver.
- 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.
- Well-articulated motivation for using Implicit Neural Representations (INR) to avoid interpolation in rotated-view EPI acquisitions, potentially improving B0 estimation accuracy over traditional blip-up/down methods.
- Experimental results demonstrate the effectiveness of INR in addressing the B0 estimation problem from rotated-view acquisitions by eliminating the need for interpolation.
- 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.
- Experiments rely on simulated k-space data with varying rotation angles instead of real acquisitions.
- Interpolation error was introduced during the generation of ground truth images in the three-view dataset, leading to degraded quantitative metrics for the proposed PINR with three-view acquisition.
- The PINR-3v result is not included in Fig. 2.
- Computational time of the proposed PINR method compared to TOPUP is not reported.
- Error maps referenced in Fig. 3 are missing.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 and promising application of INR to jointly estimate B0 inhomogeneity and reconstruct undersampled images from rotated-view EPI acquisitions, eliminating the need for interpolation. The motivation is well-articulated, with experiments showing improved performance compared to conventional methods like TOPUP. However, the study relies solely on simulated k-space data, and real acquisition results are missing, which limits the generalizability of the findings. Additionally, there are presentation issues such as missing PINR-3v results in Figure 2, absent error maps in Figure 3, and a lack of computational time comparison with baseline methods. The use of interpolation when generating ground truth for the three-view dataset may also introduce bias in the evaluation. Overall, despite these shortcomings, the core idea is innovative, the methodology is technically sound, and the results are promising.
- 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
Review #3
- Please describe the contribution of the paper
This paper proposes PINE, an unsupervised implicit neural representation-based method for addressing joint B0 estimation and image reconstruction in Echo Planar Imaging (EPI). Thanks to the continuous representation of INRs, the proposed PINE is able to capture spatial variations inherent in multi-view acquisitions. By further incorporating MRI physical models into the INR, PINE can jointly reconstruct high-quality B0 maps and MR images directly from k-space data, without using any external data. Experimental results also demonstrate the effectiveness of the proposed method.
- 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.
- By following MRI physical models, this paper proposes a new formulation (Eq. 2) for joint B0 estimation and image reconstruction in EPI, which fundamentally eliminates cumulative errors present in previous two-step methods.
- Although the application of INRs in medical imaging (e.g., MRI and CT) is well-established, its exploration in the field of joint B0 estimation and image reconstruction is novel. The use of INRs in EPI is also well motivated, in my opinion, particularly because this work effectively addresses variations in multi-view acquisitions through the continuous representation of INRs.
- This paper is well written and provides adequate background on EPI and technical details.
- 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.
In my opinion, this paper does not present any major weaknesses. I just have a few minor questions and comments.
- The objective defined in Eq. (2) is a multi-variable optimization problem. The authors use two networks to represent B0 estimation and MR images, and jointly optimize them. I wonder whether the model stability could be improved by optimizing the two networks alternatively.
- In Eq. (5), an adjoint operator is used to address the high dynamic range problem inherent in k-space data. To my knowledge, some prior works have addressed this issue. For example, Feng et al. use a related L2 loss, while Wu et al. introduce the Fourier-slice theorem to reformulate radial MRI as a back-projection problem. I think it would be beneficial to discuss the impact of these specific transformations, as this could further strengthen the reliability of the proposed method.
[1] Feng, Jie, et al. “Spatiotemporal implicit neural representation for unsupervised dynamic MRI reconstruction.” IEEE Transactions on Medical Imaging (2025). [2] Wu, Qing, et al. “Moner: Motion Correction in Undersampled Radial MRI with Unsupervised Neural Representation.” ICLR 2025.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper proposes a new formulation for addressing key challenges in EPI. By further incorporating INRs, a powerful deep learning-based solver, into the proposed optimization framework, the method shows great potential in improving reconstruction quality.
- 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
Dear Meta-Reviewer and Reviewers, Thank you for your constructive feedback and provisional acceptance. We have addressed all comments below.
Meta-Review We appreciate being among the top 9% and will complete camera-ready submission and registration.
Reviewer 1
- Limited evaluation The methods mentioned in the Introduction can correct EPI distortion but generally require additional scans. We selected TOPUP for comparison because it remains the most widely used—and state-of-the-art—EPI correction method. In a planned extension of this work, we will include a more comprehensive evaluation. We hope that the current results adequately demonstrate the effectiveness of our approach.
- Generalizability EPI and EPTI are distinct imaging modalities. EPTI leverages temporal signal evolution to produce distortion- and blur-free multi-echo images, albeit with some extra scan time, making it well suited for simulation and evaluation. EPI remains valuable for rapid acquisitions in applications such as diffusion-weighted and diffusion-tensor imaging. In our extended study, we will compare acquisition speed, image quality, and relevant application scenarios more thoroughly.
- Coarse resolution Higher resolution generally introduces greater challenges in MRI reconstruction. We have demonstrated our method at 1 mm isotropic resolution and expect similar performance at lower resolutions. We will perform more extensive resolution studies in the extended version.
- Loss typo Thank you for bringing this to our attention. We will correct Equation (5) and the accompanying text to specify a smoothed L1 loss in the camera-ready manuscript.
- Dataset description As described in Section 3, our dataset comprises three volunteers, each with 90 slices (270 slices total). We believe that restating this information in Table 1 would be redundant.
- Definition of “m” We appreciate the catch. The symbol should be “n,” representing different masks for each view; we will correct this.
- References Our references are currently ordered by the first author’s last name. We will reformat and reorder them to comply with the MICCAI style guidelines.
Reviewer 3
- Optimization stability In our experiments, joint optimization of the two networks proved stable. The rotated sampling scheme provides sufficient redundant information to converge reliably.
- Loss functions in prior work Thank you for the suggestion. Wu et al.’s loss is tailored to radial sampling trajectories and is not directly applicable to our rotated Cartesian grid. We agree that exploring alternative loss functions—such as the related L2 loss used by Feng et al.—would be a valuable direction in our extended study.
Reviewer 4
- Use of real data We acknowledge that relying solely on semi-simulated k-space data is a limitation. However, only the acquisition process (including rotated-view interpolation) was simulated; the high-resolution images and B0 maps are drawn from real samples. We believe this semi-simulated approach provides a solid initial evaluation of our methodology. We are currently extending this work to include fully real acquisitions.
- Figures We omitted the PINR-3v results from Figure 2 to focus on the direct comparison between two-view TOPUP and our method; Figure 3 then illustrates the benefits of three-view acquisition. This layout aligns with the organization of our Results section and allows sufficient space for image details. Due to space constraints, we did not include error maps in Figure 3, but the visual improvements are clear. We will add the error maps in the camera-ready version.
- Runtime Thank you for raising this point. On an NVIDIA A6000 GPU, PINR requires approximately 4 minutes per slice, whereas TOPUP processes a full volume of 90 slices in about 10 hours for this high-resolution data. We will include these timings in the Experimental Setup section of the camera-ready manuscript.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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