List of Papers Browse by Subject Areas Author List
Abstract
Implicit neural representations (INRs) have demonstrated significant promise as a patient-specific alternative to data-driven deep learning (DL) for accelerated MRI reconstruction. However, without additional priors, they struggle to match state-of-the-art supervised reconstruction performance, particularly at higher acceleration factors. To address this, recent research has explored leveraging minimally biased information sharing between multiple contrasts within a subject’s own dataset. In this work, we demonstrate that straightforward extensions of INR methods to the multi-contrast setting do not efficiently capitalize on the complementary information present in the shared underlying structures between contrasts. We therefore propose a novel structure-contrast DISentangled Implicit Neural Representation (DISINR) framework for multi-contrast MRI reconstruction, specifically designed to disentangle the shared structural information from each contrast image during the training process. We demonstrate that our method out-performs previous patient-specific multi-contrast and single-contrast methods on the 0.3T M4RAW dataset, producing robust reconstructions at accelerations up to 8x. Additionally, we visualize the internal features of our model and demonstrate that it learns a disentangled structure representation, enabling it to better support mixed acceleration protocols. Our code is available at: https://github.com/chiew-group/DISINR.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0094_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/chiew-group/DISINR
Link to the Dataset(s)
M4RAW dataset: https://github.com/mylyu/M4Raw
BibTex
@InProceedings{VavZac_StructureContrast_MICCAI2026,
author = { Vavasour, Zach AND Vavasour, Kyle AND Chiew, Mark},
title = { { Structure-Contrast Disentangled INRs for Accelerated Multi-contrast MRI Reconstruction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16888},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes DISINR, a structure-contrast disentangled implicit neural representation framework for patient-specific multi-contrast accelerated MRI reconstruction. The method introduces a disentangled encoder-decoder INR architecture with structure feature masking to better leverage shared anatomical information across contrasts during joint reconstruction.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The paper addresses a well-motivated and practically relevant problem in patient-specific multi-contrast accelerated MRI reconstruction.
- The proposed disentangled INR architecture is technically sound and specifically tailored to encourage structural information sharing across contrasts.
- The experimental design is solid, with appropriate comparisons against both single-contrast and multi-contrast patient-specific reconstruction baselines, as well as ablation studies validating key 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.
- The quantitative improvements over the strongest baseline are relatively modest. The gains over MC-IMJENSE are limited to approximately 0.2–0.3 dB PSNR with minimal SSIM improvement, and the qualitative visual differences appear subtle.
- Evaluation is limited to a single low-field dataset. All experiments are conducted on the 0.3T M4RAW dataset, leaving generalizability to conventional higher-field or more diverse clinical MRI acquisitions unclear.
- The method assumes well-aligned multi-contrast acquisitions. The proposed disentanglement framework may be sensitive to inter-contrast misalignment, which could limit applicability in more realistic clinical settings.
- 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.
- 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?
Overall, this paper presents a well-motivated and technically sound approach for patient-specific multi-contrast INR reconstruction. While the quantitative improvements over prior multi-contrast INR methods are modest, the proposed disentangled formulation is thoughtfully designed for the problem setting and supported by appropriate ablations. Despite limitations in evaluation breadth and practical gain magnitude, I believe the paper offers a meaningful incremental contribution to the growing literature on patient-specific MRI reconstruction and would support 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
Review #2
- Please describe the contribution of the paper
This paper proposes DISINR, an unsupervised method based on implicit neural representation (INR) for multi-contrast MRI reconstruction. Its key insight is that MR images with different contrasts from a single subject share the same anatomical structures. This observation motivates the introduction of contrast-specific and structure-shared encoders, which disentangle contrast and structural features. This work compares the proposed DISINR with IMJENSE, a state-of-the-art INR-based method, and also conducts ablation studies on DISINR. The empirical results confirm the effectiveness of the proposed method. Overall, in my opinion, the motivation and idea of this work are novel, and the paper is suitable for publication at MICCAI.
- 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.This work leverages the key insight that MR images with diverse contrasts from a single subject share the same anatomical structures. This motivation and idea are reasonable and novel. 2.The introduction of a contrast-specific encoder to decompose contrast information is meaningful and helps improve MRI reconstruction performance. 3.This work conducts preliminary experiments that demonstrate the effectiveness of the proposed method. 4.Disentangling shared and contrast-specific features is a good idea. I think it may inspire more applications in the field of INR-based MRI reconstruction.
- 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.In the proposed DISINR, the shared structural feature is pixel-wise and represented by a continuous function of coordinates. However, the contrast feature is generated only by a linear projection layer from a contrast index. In other words, the contrast feature is global (e.g., for T1w, the contrast information is identical for the whole image). In my humble opinion, this contrast representation is too compact. Why not also use a continuous function to model it? 2.All experiments are conducted on a low-field 0.3T dataset. I would like to know the reason for choosing this low-field dataset. In practice, there are many MRI datasets for evaluating MRI reconstruction, such as fastMRI. Does the proposed DISINR also work on these higher-field datasets? 3.The improvement of DISINR over the existing IMJENSE is somewhat limited (less than 0.5 dB).
- 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.
- 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?
See strengths and weaknesses, please.
- 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
The paper proposed a structure-contrast disentangled implicit neural representations (INR) framework for accelerated multi-contrast MRI reconstruction. The key idea is to separate shared anatomical structure from contrast-specific information. The main contribution of the paper is the introduction of dual-encoder framework consisting of a structure encoder that maps spatial coordinates to a structural latent representation, and a contrast encoder that maps a contrast index to a contrast latent. The contrast encoder acts as a prior to the structural encoder and their latent representations are concatenated and fed into the decoder to reconstruct contrast-specific images.
- 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 paper addresses an important problem in accelerated MRI reconstruction and the idea is clearly presented. Multi-contrast MRI is often acquired jointly, and leveraging shared structural information is well justified for undersampled reconstruction. Disentangling multi-contrast information to extract the anatomical structure is intuitive and clearly explained. 2.The introduction of a dedicated contrast encoder combined with a structure encoder is a meaningful contribution. The additional use of random masking on structure features is particularly interesting, as it explicitly encourages the network to avoid encoding contrast information in the structure encoder. The ablation study demonstrates the structure feature masking plays an important role in performance improvement, supporting the authors’ design choices. 3.The evaluation of the network on low-field (0.3T MRI) is a strength, as it represents a challenging and clinically relevant scenario. Demonstrating robustness under low-SNR conditions adds practical value to the work.
- 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.While the proposed framework is well designed, the core idea of disentangling shared structure and contrast is not entirely new. Prior works have explored similar concepts in multi-contrast MRI reconstruction and representation learning, e.g. Niessen et al. [12], NeRP [14] and IMJENSE [5] also leveraged shared representations across contrasts. The paper would benefit from a clearer positioning of what is fundamentally novel beyond adapting disentanglement into the INR framework. 2.The quantitative improvements over prior methods are relatively modest (e.g., ~0.2–0.3 dB PSNR gains and ~0.005-0.007 SSIM gains over MC-IMJENSE as shown in Table 1). While consistent, these gains are small and raise questions about practical significance, especially given the added architectural complexity. 3.The paper reports mean and standard deviation metrics but does not include statistical significance testing. Given the small performance margins, such analysis is important to validate whether the improvements are meaningful.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if 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 addresses an important problem and proposes a well-motivated framework, including the contrast encoder and structure feature masking. The method is clearly presented and supported by ablation studies. However, the main factors leading to my decision are the limited novelty relative to prior works on multi-contrast MRI and INR-based methods, and the modest qualitative and quantitative improvements over existing methods. The reported gains are small and not supported by statistical significance analysis, making it unclear whether they are meaningful in practice.
- 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 my comments, and I have no further major concerns.
Author Feedback
We thank reviewers and meta-reviewer for their valuable feedback. We have summarized the points and provide our response as follows.
1.R3 views the core idea of structure / contrast disentanglement as not entirely new. To the best of our knowledge, Niessen et al. [12] is the only MRI reconstruction method that jointly models multiple contrasts with an INR, and this work did not explore structure/contrast disentanglement. R3 views NeRP [14] and IMJENSE [5] as having leveraged shared representation across contrasts; however, both methods have only explored single contrast reconstruction. We believe that our proposed framework for inter-contrast information sharing remains an entirely new concept in the field of zero-shot reconstruction.
2.R1,R2 view the evaluation solely on the M4RAW dataset as a weakness. To the best of our knowledge, M4RAW is the only publicly available raw MRI dataset consistently collected with paired contrasts. R2 proposes fastMRI, and while subsets of the dataset do contain subjects with multiple contrasts, they are inconsistent and exhibit mismatched resolution, alignment, and matrix sizes, making them not directly suitable for multi-contrast INR reconstruction. We note that because our method is subject-specific (i.e. zero-shot) we do not face the same domain-shift challenges common to supervised methods; therefore, we strongly believe that our method will generalize well to clinical field strengths.
3.R1 comments that our method may be sensitive to inter-contrast misalignment. We do not view this limitation to be insurmountable within our proposed framework. We expect that confounds such as motion can be addressed by transforming the co-ordinate inputs to align the internal continuous representations without any other modifications to the image models. Leveraging this, we are actively working on extending our method to handle misalignment and other sources of heterogeneity that would be encountered in real-world settings.
4.R2 suggests that the contrast branch of the network is too compact, and could be replaced with something more expressive. Our fundamental hypothesis is that the majority of information within a set of matched contrast images is contained within the shared anatomical structure, and that the contrast differences relative to the shared structure can be encoded with far less complexity; therefore, we believe that the limited expressivity of the contrast branch is key for our proposed structure/contrast separation. However, it may be interesting to explore contrast encoder complexity in the future.
5.R1,R2,R3 view the quantitative improvements as modest, additionally R3 notes that no statistical significance testing was performed. Firstly, we can confirm that the originally presented PSNR/SSIM improvements of our method relative to the next best performing model MC-IMJENSE are statistically significant across all contrasts and acceleration factors (p < 1e-4 in all cases). Additionally, we note that the M4RAW dataset represents an extremely noisy reconstruction context, where the baseline quality of the data likely confounds efficient information sharing between contrasts. Given this, we believe that the consistent performance gains of 0.2-0.3dB across all tested acceleration rates demonstrates the strong theoretical basis for our idea of structure contrast disentanglement.
Additionally, we wish to correct a mistake that was made in our proposed interpretation of the learned contrast embeddings (Fig. 4 b). We originally claimed that they are generalizable between subjects despite entirely independent and uniquely initialized training runs. This is inaccurate, as the existing figure was made with identically-initialized training runs. With full random initialization, the contrast embeddings are not aligned across subjects, as each subject specific training run learns the structure/contrast separation in a unique way. An updated figure would be included in the final paper.
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 proposes an INR-based framework for accelerated multi-contrast MRI reconstruction that disentangles a shared structural latent from a contrast-specific latent space. The reviewers agree on several strengths: the problem is well motivated, the architecture is technically sound, the structure feature masking is an interesting idea supported by ablations, and the experimental protocol includes both single- and multi-contrast baselines.
The AC found reviewers split (4/4/3) on three main points. The first is novelty. R2 views the motivation and idea as novel, whereas R3 argues that shared structure/contrast representations have already been explored in multi-contrast MRI reconstruction (citing Niessen et al., NeRP, and IMJENSE) and that the paper needs clearer positioning of what is fundamentally new beyond porting disentanglement into INR. R1 takes a middle position, treating the contribution as incremental but meaningful. The second concern is whether the modest quantitative gains (~0.2–0.3 dB PSNR, ~0.005–0.007 SSIM over MC-IMJENSE) are sufficient; R1 and R2 accept them as part of an incremental contribution, while R3 views them as too small to justify acceptance without statistical significance testing. The third weakness is the sole use of the 0.3T M4RAW dataset, which R3 frames as a strength (low-field, clinically relevant) but R1 and R2 view as a limitation on generalizability.
The AC views these points primarily clarification issues, rather than fundamental flaws in the experiments/comparisons. Given the split scores and the tractable nature of the concerns, I recommend inviting the authors to respond in rebuttal before a final decision. Please carefully read reviewers’ comments and MICCAI rebuttal guidelines while preparing the rebuttal.
- 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.
This paper proposes a structure-contrast disentangled INR for multi-contrast MRI reconstruction, and while the quantitative gains over MC-IMJENSE are admittedly modest, the rebuttal confirmed they’re statistically significant across all settings and gave a reasonable justification for the single-dataset evaluation, with R3 moving to accept and R1 holding a positive score. The core idea is interesting and the disentanglement design is supported by ablations, so I’m recommending acceptance. The camera-ready should include the corrected Fig. 4b interpretation the authors acknowledged during rebuttal.
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.
The authors have addressed the reviewers’ concerns in their rebuttal. Specifically, they have addressed the novelty of the method and provided statistical tests to show that, while modest, the improvements described are statistically significant. I believe that this paper represents a solid contribution that should be recommended 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.
Accept
- Please justify your recommendation.
The authors clarified their novelty and well addressed the concerns of R3 who rated the submission as weak reject before rebuttal. I vote for accept and feel this is a timely paper using INR and enforcing structural disentenglement. It would raise discussion at MICCAI 2026.