List of Papers Browse by Subject Areas Author List
Abstract
As an advanced MRI technique, multi-parametric magnetic resonance imaging provides valuable quantitative maps but necessitates long scan times. Accelerated imaging techniques are therefore essential, yet existing reconstruction methods often deliver suboptimal performance due to cascading errors between the reconstruction and quantification steps, and furthermore, lack adequate physical constraints on the signal model. To address these challenges, we propose a Physical-Driven Unified Implicit Regularization Network (PDUIR-Net) for joint estimation of images and quantitative maps in MRI. Specifically, the network is derived by unfolding the solver of an implicit-regularization reconstruction model with a unified objective. We further design a prompt-guided decoupled reconstruction module to disentangle shared and contrast-specific features across different contrasts, and develop a proximal quantitative mapping module to learn priors directly from data. Consequently, all components are integrated into an end-to-end trainable framework. Experimental results on our in-house data (12-echo sequence) show that our method outperforms the state-of-the-art methods by 1.18\% in SSIM for multi-contrast images, and 0.89\% in SSIM for quantitative maps at 10× acceleration, which significantly advances the performance limitation for multi-parametric MRI.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0756_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/xjtuMIMILAB/PDUIR-Net
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YanYan_PhysicalDriven_MICCAI2026,
author = { Yang, Yan AND Xiao, Han AND Yang, Xue AND Wang, Jiazhen AND Jia, Sen AND Zheng, Hairong AND Sun, Jian},
title = { { Physical-Driven Unified Implicit Regularization Network with Scan Parameter Prompts for Accelerated Multi-parametric MR Imaging } },
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 introduces PDUIR-NET, a physics-driven deep network designed for the joint estimation of multi-contrast images and quantitative maps in MRI. The main contributions of this work are: (1) proposing a novel physics-driven unrolled network architecture for joint reconstruction and quantification; (2) utilizing scan parameters as prompts to adaptively modulate a Prompt-Guided Decoupled Reconstruction Module; and (3) learning a prior on signal ratios across multi-contrast images to indirectly model quantitative maps.
- 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.
Use of Prompts: Using scan parameters as prompts to adaptively modulate the network allows improving the final image quality
Mathematical Foundation: Employing the Half-Quadratic Splitting algorithm to optimize the physics-based model provides a theoretical grounding for the unrolled network design, bridging the gap between deep learning and inverse problems.
Ablation Study: The paper includes an ablation study that systematically evaluates the effectiveness of the three key components of PDUIR-Net, successfully verifying the necessity of each proposed module.
- 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.
Presentation and Clarity Issues: The manuscript suffers from several organizational and typographical issues that hinder readability. For instance, the caption of Figure 1 refers to sub-panels (a), (b), and (c), but these labels are missing from the figure itself. Furthermore, Equation 1 introduces mathematical notations such as ‘A’ (forward operator) and ‘G(X)’ (regularization term). Although these variables are briefly defined alongside the equation, detailed explanations regarding their specific formulation or implementation are lacking. The authors should provide these missing details to ensure full clarity and reproducibility.
Missing Details in the Ablation Study: While the inclusion of the ablation study is a strength, the manuscript lacks methodological details on its execution. Specifically, when a core component is eliminated, there is no explanation of how the network handles the altered inputs and outputs, or how the architecture is adjusted to ensure a fair comparison.
- 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 novel and promising conceptual framework (PDUIR-NET) for the joint estimation of MRI images and quantitative maps. The use of scan parameters as prompts and the integration of a signal ratio prior are particularly interesting methodological contributions.
The current manuscript suffers from several presentation issues and a lack of detailed explanations in certain areas. Despite these unpolished aspects, the core conceptual contributions are strong and valuable to the MRI reconstruction community.
- 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 rebuttal directly and effectively addresses my two main concerns.
First, on presentation and notation, the authors clarify the missing definitions in Eq. (1). Second, on the ablation study, the authors provide a clear row-by-row description of how each ablated variant handles its altered inputs and outputs.
Given that the core conceptual contributions are valuable, my original concerns were primarily about clarity and ablation interpretability, and these are now resolved, I maintain my recommendation.
Review #2
- Please describe the contribution of the paper
The main contribution of this paper is the introduction of an end-to-end unrolled reconstruction framework for accelerated multi-parametric MRI that jointly estimates multi-contrast images and quantitative maps. Instead of treating image reconstruction and quantitative mapping as two separate stages, the method couples them through an alternating optimization scheme and unfolds this scheme into a multi-stage network. Within this framework, the paper further designs dedicated modules to incorporate physical consistency across contrasts, exploit shared and contrast-specific priors, and improve parameter efficiency through scan-parameter-guided adaptive blocks. Overall, the paper’s central contribution lies in integrating reconstruction and quantitative estimation into a single learnable architecture informed by MRI physics and inter-contrast relationships.
- 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 paper proposes a unified framework for jointly reconstructing multi-contrast images and estimating quantitative maps, which is more principled than a conventional two-stage pipeline.
2.The architecture is also well designed: it explicitly models shared and contrast-specific information across contrasts, and introduces adaptive parameter-sharing mechanisms to improve efficiency and reduce redundancy. and the method is physics-informed and optimization-inspired, rather than being a purely black-box network.
3.The experiments report consistent, albeit modest, improvements across two acceleration factors, and the ablation study provides some support for the usefulness of the proposed components.
- 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 presentation of the method contains some notation and consistency issues that make the technical description harder to follow. For example, 𝑃 is introduced before Eq. (1) as the number of pixels, but in Eq. (2) it appears to denote the number of contrast pairs, which creates unnecessary ambiguity. In addition, in the PQMM subsection, the update of 𝑅 and 𝑆 is described as corresponding to Eq. (3c), whereas from the formulation it should instead correspond to Eq. (3b).
2.The claimed benefit of the prompt-guided design is not yet fully substantiated. The prompts in this work are acquisition parameters from a relatively narrow and fixed protocol, and it remains unclear how well this conditioning strategy would generalize to more diverse acquisition settings, scanner vendors, or protocol variations. Since quantitative MRI protocols are often flexible in practice and can vary substantially across clinical environments, the robustness of the proposed prompt-guided mechanism beyond the specific in-house setting is currently not demonstrated.
3.The study includes only 33 subjects in total, with 24 used for training, 4 for validation, and 5 for testing. Given that the method is a multi-stage 3D unrolled network with several dedicated components, this sample size raises a legitimate concern about overfitting and limits the strength of the generalization claims. At present, the experimental evidence is not enough as a strong validation of this work.
4.The experimental section lacks sufficiently strong statistical support. Only four subjects are used for validation, and the paper reports no statistical analysis such as standard deviations, confidence intervals, or significance testing. This is particularly important because the reported improvements over competing methods and in the ablation study appear relatively modest. With such a small evaluation set, it is difficult to determine whether the observed gains are consistently meaningful or within the range of subject-level variation.
5.The paper states that the model uses coil-combined data with sensitivity maps computed by ESPIRiT, but it is not sufficiently clear how this setup would translate to a more realistic accelerated acquisition scenario, especially under highly undersampled conditions. The manuscript should better explain the calibration assumptions and discuss how robust the pipeline would be in practical prospective settings.
6.Reproducibility is currently weak. The study is based on an in-house dataset and does not provide open-source code, and several implementation details that are important for reproducing the method are missing. For example, the exact design of the prompt embedding and the detailed architecture of the adaptive modules. This makes it difficult for others to verify or build upon the work.
7, The paper argues that the architecture is parameter-efficient, but it does not provide quantitative evidence to support this claim. Reporting model size, training time, inference time, or memory usage would make the efficiency argument much more convincing, especially since the method introduces multiple modules and is trained on large 3D multi-contrast data.
- 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
Given the complexity of the proposed method, releasing code and a minimal demo would greatly improve the paper’s transparency and reproducibility, especially under MICCAI page constraints. In addition, the current dataset size seems limited relative to the complexity of the model, and stronger statistical validation would be needed to better support the significance and robustness of the reported improvements.
- 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?
On the positive side, the manuscript proposes an interesting unified framework for multi-contrast reconstruction and quantitative mapping, and the modeling of shared and contrast-specific information is thoughtful. The overall method is technically appealing and shows potential.
However, several factors limit my confidence in the work in its current form. First, the prompt-guided design is central to the paper, yet its robustness and generalizability beyond the relatively fixed acquisition setting used in this study remain unclear. This raises concerns about how well the method would transfer to broader clinical or cross-protocol scenarios. Second, a number of important implementation details are insufficiently described, including aspects of the prompt embedding and the architectural design of key modules, which makes the work difficult to fully assess and reproduce. Third, the experimental validation is based on a relatively small in-house dataset, and the reported improvements are modest, without sufficient statistical analysis to establish their significance. Taken together, these issues reduce the strength of the empirical evidence relative to the ambition of the proposed method.
- Reviewer confidence
Very confident (4)
- [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 primary concern is that the current experimental evidence is obviously insufficient to support the paper’s key claimed innovation, namely the prompt-guided adaptation using scan parameters. In the current study, the prompts (TR, TE, FA) appear to be tied to a relatively fixed MTP protocol that defines the 12 contrasts. Under this setting, it is difficult to determine whether the prompts truly enable meaningful scan-parameter-guided adaptation, or whether they mainly act as fixed identifiers of the predefined contrasts. Also the reported improvements over the baselines are relatively modest, the current results do not yet convincingly demonstrate that the prompt design is the main source of gain, nor do they establish its robustness under broader protocol variability.
A second major concern is that the rebuttal introduces a new source of ambiguity regarding the actual implementation. The manuscript describes undersampling on the kz plane of a 336×288×84 acquisition, which strongly suggests a 3D acquisition setting. In contrast, the rebuttal states that the proposed method is a 2D model trained on 6,760 slices. This does not necessarily imply that the method is incorrect, but it does leave the implementation insufficiently specified. If the acquisition and forward model are 3D, while the network is applied slice-wise in 2D, the manuscript should explicitly explain how the forward operator is handled, and whether the reconstruction is 3D or a 2D approximation. At present, the methodological description and the experimental implementation do not align clearly enough, which makes the work difficult to interpret and assess.
My third concern remains the limited scale of validation. Although the rebuttal clarifies that the model is trained slice-wise, the effective diversity is still constrained by the small number of subjects and the single in-house protocol. The additional generalization claim (training on 10× Poisson-disc and testing on 9× CAIPI / prospective 9× data) is encouraging, but in its current form it is not yet sufficient to establish robust generalization.
Overall, while the rebuttal resolves several minor presentation issues, it does not fully address my central concern that the core prompt-based contribution and the actual reconstruction setting remain insufficiently validated.
Review #3
- Please describe the contribution of the paper
This paper proposes a physics-driven unrolled network for joint multi-contrast reconstruction and quantitative mapping, integrating signal-ratio-based physical constraints and scan-parameter prompts into a unified implicit regularization framework.
- 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 proposes a well-motivated unified framework for joint multi-contrast reconstruction and quantitative mapping, integrating data fidelity, signal physics, and learned priors. The use of signal-ratio-based physical constraints provides a practical way to incorporate MR physics while maintaining tractability. The introduction of scan-parameter prompts is a novel and effective design for adapting the network across contrasts. The method is systematically derived via unrolling, improving interpretability, and is validated on a reasonably sized in-house dataset with consistent gains over strong baselines.
- 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 overall formulation is largely incremental over existing physics-driven unrolled frameworks for joint reconstruction and mapping (e.g., MANTIS, SRM-Net, and other HQS/ADMM-based unrolled models), and many components such as proximal networks, U-Net mapping modules, and alternating optimization are standard. The use of scan parameters as conditioning (prompts) is related to prior conditional or sequence-aware MRI networks, limiting its novelty. In addition, the signal-ratio-based physical modeling, while practical, is heuristic and only approximates underlying Bloch dynamics without clear analysis of its accuracy. Experimentally, the evaluation is limited to a single in-house dataset with no external validation, and comparisons to more recent strong baselines (e.g., diffusion-based or advanced physics-informed methods) are missing, making it difficult to assess generalization and real-world impact.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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 presents a well-motivated physics-driven framework for joint multi-contrast reconstruction and quantitative mapping, with a unified formulation that integrates data fidelity, signal modeling, and learned priors. The use of signal-ratio-based constraints and scan-parameter prompts is practical and improves adaptability across contrasts, and the unrolled design provides reasonable interpretability. The method is validated on a moderately sized in-house dataset and shows consistent improvements over several strong baselines.
However, the overall novelty is somewhat incremental, as the framework largely builds on existing unrolled optimization approaches (e.g., MANTIS, SRM-Net) with standard components. Some design choices, such as the signal-ratio modeling, are heuristic and lack deeper analysis. In addition, evaluation is limited to a single dataset without external validation or comparison to more recent methods (e.g., diffusion-based or stronger physics-informed approaches). Overall, the paper is technically sound with solid engineering, but the level of novelty places it slightly above the acceptance threshold.
- 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.
Although some concerns remain, including incremental novelty, limited analysis of the signal-ratio approximation, lack of external validation, and missing stronger recent baselines, I still find the work technically sound and potentially useful. The proposed framework is well motivated, integrates data fidelity, signal modeling, and learned priors in a coherent way, and shows consistent improvement over the evaluated baselines. Since my original assessment was slightly above the acceptance threshold and the rebuttal did not reveal any critical flaw, I maintain an Accept recommendation.
Author Feedback
We thank reviewers for their insightful comments. All clarifications in this rebuttal will be in the revised paper. [R1 & R2-Q1] Textual errors. Missing sub-figure labels (a)-(c) in Fig. 1 are added. In Eq. (1), A=MFS denotes a MRI forward operator, with M, F, and S representing the sampling mask, Fourier transform, and coil sensitivity, respectively; G(X) is a data-driven prior for multi-contrast images; D is the pixel number. In Eq. (2), P is the number of image pairs for computing R. In the PQMM subsection, Eq. (3c) is corrected to Eq. (3b). [R1] Ablation study implementation details. In Table 2, Row 4 is the full proposed network. Row 3 removes PCM: it uses the network to reconstruct multi-contrast images only, with final quantitative maps obtained via traditional estimation. Compared to Row 3, Row 2 removes AUB-G and uses independent common/unique filters (Q_i^c, Q_i^u) per contrast, without scan-parameter prompts to modulate shared filters. Row 1 further removes AUB-U and uses contrast-specific update networks Net_Ui instead of a shared adaptive update network Net_U. [R3] Novelty and Contributions. Unlike existing unfolding methods, we address quantitative task-driven MRI reconstruction for MTP, where image reconstruction serves as an intermediate step for accurate parameter estimation. We design a new optimization objective integrating MRI signal physics with data acquisition mechanism, explicitly decoupling shared and contrast-specific features for multi-contrasts, and incorporating quantitative-driven implicit priors. Unlike existing multi-contrast reconstruction methods, we introduce core scan parameters (TR, TE, FA) as prompts to modulate contrast-specific filters, enabling a single shared adaptive network across contrasts for better parameter and computational efficiency. Contrary to quantitative methods like MANTIS/SRM-Net, we estimate signal ratios instead of direct parameters, converting nonlinear mappings into tractable linear constraints to avoid gradient instability from complex exponential/division operations in direct T1/T2* estimation. Finally, we adopt alternating optimization between quantitative maps and images, letting maps serve as a physical common prior that inversely constrains image reconstruction. [R2-Q2,Q7] Scan Parameter Capability and Network Efficiency. As TR, TE, and FA determine the 12 MTP contrasts, we use them as prompts to modulate filters in a single shared network for intra-sequence contrast adaptation. This avoids per-contrast separate networks, cutting parameters from 50M to 10M and training time from 80 to 60 hours. [R3 & R2-Q3,Q4,Q6] Datasets Scarcity and Generalization Capability. MTP is an advanced proprietary technology with no public dataset. Full-sampling data is costly and difficult to obtain, limiting current subject number. Nevertheless, it is worth noting that our method is a 2D model trained on 6,760 slices from 33 subjects, offering sufficient coverage. Its generalization is supported by two aspects: a lightweight architecture with embedded MRI physics priors to reduce data dependency, and validation under unseen acquisition settings. We trained the network with a 10× Poisson-disc mask and directly tested it on 9× CAIPI sampling data, achieving an NRMSE of 0.064.We also tested the network on prospectively acquired 9× clinical data, which showed clear details, demonstrating its robustness in real scenarios. Upon acceptance, we will release the code and dataset. To our knowledge, no existing diffusion-based method directly addresses this specific problem. We will explore a diffusion-based method by incorporating MRI physics and multi-contrast signal priors in future work. [R2-Q5] Practical Utilization of Coil Sensitivities. MTP scans all contrasts simultaneously, so they share the same coil sensitivity maps. These maps are estimated from a low-resolution 24×24×24 pre-scan taking only 10 seconds, making sensitivity estimation feasible even at high acceleration factors.
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.
Strengths: The paper proposes a well-motivated physics-driven unified framework for joint multi-contrast MRI reconstruction and quantitative mapping, with scan-parameter prompts, shared/contrast-specific modeling, and ablation evidence supporting its main modules. Weaknesses: Reviewers were concerned that the novelty is somewhat incremental over existing unrolled physics-based methods, the presentation has notation and clarity issues, and the evidence is limited by a small in-house dataset, weak statistical support, uncertain prompt generalizability, and incomplete reproducibility details. The authors therefore be invited for 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.
The post-rebuttal assessment leans more toward Accept than Reject. Nevertheless, the authors are advised to address the reviewers’ remaining concerns.
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.
This paper proposes PDUIR-Net, a physics-driven unified implicit regularization network with scan-parameter prompts, to jointly reconstruct multi-contrast MR images and estimate quantitative maps for accelerated multi-parametric MRI. Reviewers R1 and R3 recognized the paper’s well-motivated unified framework, its integration of data fidelity, signal modeling, and learned priors, as well as the use of scan-parameter prompts and signal-ratio-based physical constraints. However, R2 remained concerned that the current experimental evidence is insufficient to support the key claimed innovation, namely prompt-guided adaptation using scan parameters. In particular, the prompts appear to be tied to a relatively fixed MTP protocol, making it unclear whether they provide meaningful scan-parameter-guided adaptation or mainly act as fixed identifiers of predefined contrasts. R2 also raised concerns about the modest improvements, limited validation on a small in-house dataset, and ambiguity between the 3D acquisition/forward model and the 2D slice-wise implementation. Although the rebuttal clarified several presentation and implementation details, it did not effectively address R2’s central concerns regarding the core prompt-based contribution, the actual reconstruction setting, and the strength of the validation. Therefore, I recommend rejection.
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.
My recommendation of Accept is based on a strong and clear methodology contribution. The rebuttal addressed concerns from reviewers 1 and 3.While there are valid concerns from Reviewer 2 regarding utility and clarity on experiments, I find those are minor.
Meta-review #4
- 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 rebuttal addressed several clarity and implementation-detail concerns and added useful information on model size, training time, coil sensitivity estimation, and testing under a different sampling pattern/prospective setting. However, the manuscript should be revised carefully to make the 2D/3D reconstruction setting explicit, clarify the forward model and undersampling assumptions, and avoid overstating generalization of the prompt mechanism beyond the evaluated protocol. The final version should also clearly incorporate the rebuttal clarifications on notation, ablation variants, parameter efficiency, and data/code release plans.