List of Papers Browse by Subject Areas Author List
Abstract
Patient motion during magnetic resonance imaging (MRI) acquisition leads to severe artifacts that degrade diagnostic quality. End-to-end deep learning (DL)-based motion correction methods are prone to structural hallucinations and often exhibit limited generalizability across different image contrasts. While some studies instead estimate motion parameters and perform model-based reconstruction to avoid hallucinations, they are typically restricted to specific k-space sampling patterns. To address these challenges, we present a novel retrospective motion correction framework that synergistically integrates DL-based motion estimation with model-based optimization and reconstruction. Our hybrid approach consists of two key components: (1) A pre-trained neural network that explicitly processes both undersampled k-space data and sampling patterns to infer motion parameters, enabling flexible adaption to variable numbers of shots and phase-encoding lines; specifically, this network combines convolutional neural network with gated recurrent unit, and is trained on a diverse simulated dataset encompassing varying image contrasts, anatomical orientations, and sampling patterns; (2) An instance-wise autofocusing algorithm that jointly refines the initial motion parameters and updates coil sensitivity maps, ultimately yielding artifact-free images. The proposed framework generalizes well across different image contrasts, k-space sampling patterns, and acceleration factors. Validations on fastMRI dataset and in-house MRI data demonstrate that our approach outperforms state-of-the-art methods in motion artifact suppression, while preserving strict physical fidelity and strong cross-protocol generalization capability. These results highlight the potential of our approach for robust clinical deployment.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3249_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiuWei_SAMoCo_MICCAI2026,
author = { Liu, Weijia AND Tang, Runyu AND Yao, Ye AND Li, Yao AND Shan, Hongming AND Wei, Hongjiang AND Jiang, Dengrong},
title = { { SAMoCo: A Learning-Initialized and Physics-Refined Framework for Sampling-Agnostic Motion Correction in MRI } },
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 presents SAMoCo, a hybrid framework for retrospective MRI motion correction that combines learning-based motion initialization with physics-based refinement. Specifically, the method first uses a sampling-pattern-aware network to estimate shot-wise motion parameters from undersampled k-space, and then refines these parameters jointly with coil sensitivity maps through an autofocusing-based model-driven reconstruction process. Experiments on fastMRI and in-house data show consistent improvements over several representative baselines under different acceleration settings.
- 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 sampling mask is used as a dynamic routing mechanism to extract per-shot PE lines before Bi-GRU temporal modeling, allowing a single trained network to generalize across varying shots, ACS lines, and acceleration factors without retraining — directly addressing the fixed-trajectory limitation of prior methods. 2.Most existing methods assume perfectly known sensitivity maps, which is unrealistic under simultaneous undersampling and motion corruption. SAMoCo incorporates sensitivity estimation into an alternating inner-outer optimization loop alongside motion refinement, treating the two as a principled coupled estimation problem. 3.The ablation quantifies motion parameter MAE directly (>70% reduction after autofocusing), disentangling each component’s contribution beyond image metrics alone. Testing on unseen FLAIR contrast and a different-vendor scanner further supports cross-contrast and cross-site generalization.
- 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.
Major Concerns: 1.It is unclear how well the motion estimation network generalizes beyond the training shot-number distribution. Since the method emphasizes sampling-agnostic motion estimation, it would be important to evaluate whether the network remains stable when the number of shots at test time exceeds or differs from the training range. 2.The sensitivity of the full pipeline to initialization quality is not sufficiently analyzed. Since the method is explicitly learning-initialized and physics-refined, the paper should better clarify how robust the final reconstruction is to different qualities of initial motion estimates. 3.The contribution of joint sensitivity-map updating is not sufficiently isolated. In particular, it would be helpful to show how much performance is lost if the method refines only motion parameters while keeping the sensitivity maps fixed. Minor Concerns: 4.The paper should describe the sampling-pattern-guided shot-wise feature extraction more clearly. In particular, the aggregation operation used to form shot-wise feature vectors is not sufficiently specified. 5.The computational cost of the iterative refinement is not clearly reported. Since the method relies on inner/outer optimization loops, reporting the number of iterations, convergence behavior, and runtime overhead relative to the baselines would improve the practical assessment. 6.The motion-correction evaluation remains relatively limited in scale, especially given the strong claims on generalization.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The work is physically grounded, the ablation is more rigorous than typical, and the cross-site generalization results are encouraging. However, each individual component has clear precedent and the contribution is primarily their combination, which reads as incremental. The evaluation is modest (60 slices, 6 subjects), all motion is simulated with no real-motion validation, inference time is never reported, and the sensitivity map modeling choice is neither ablated nor compared against standard alternatives such as ESPIRiT.
- 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 work presents a retrospective motion correction framework that is a hybrid of deep learning and model-based methods. A deep learning model estimates motion parameters from undersampled k-space data and k-space sampling patterns, and the model-based algorithm and framework update coil-sensitivity maps and reconstruct artifact-free images. This method demonstrates SOTA performance on one publicly available and one in-house datasets.
- 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 work is relevant, well-motivated and well-organized, and free of grammatical errors and typos
- Figures are neat and legible
- 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 motion simulation method is not explained in sufficient detail.
- Non-cartesian sampling patterns are not considered, and no justification for exclusion is presented.
- Robustness of the method on pathological data is not evaluated.
- Downstream performance evaluations are not considered. For example, does segmentation performance on the motion-corrected data approach that of the ground truth?
- 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
SAMoCo is never expanded anywhere in the text
- 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?
While the work is relevant and presented clearly, it is not easily reproducible. Further, the scope of the work is limited to cartesian trajectories and healthy data, making it difficult to assess the value and impact of the motion-correction method in real-world settings.
- 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 #3
- Please describe the contribution of the paper
A hybrid retrospective motion-correction framework is proposed that performs a DL–based motion estimation coupled with a physics-driven optimization and reconstruction framework. The DL model is pre-trained to estimate rigid body motion parameters such as translation and rotation from each k-space shots. A pre-training stage is utilized where rigid motion is simulated with random motion parameters for multiple contrasts, anatomical orientations etc. to predict motion parameters from k-space shots. The pre-trained network is simply used for initialization after which it has no role. An iterative model based reconstruction (physics driven autofocus method) is used to jointly refine the motion parameters and estimate the coil sensitivity maps. The outer loop updates the coil sensitivities and the inner loop updates the motion parameters. The method is tested using retrospective under sampling experiments and simulated motion on IXI datasets, fastMRI datasets and in-house datasets. The method is compared to AF+ and JSMoCo in addition to other baseline methods with no motion compensation and is shown to outperform those methods at acceleration factors of 2 and 3.
- 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 DL-part is only used as an initialization for motion parameter estimation. The motion corrected reconstruction is performed using physics-based formulations such as autofocus. Thus there is no hallucination risk, which could be a real risk if it were an end-to-end DL approach. The DL-network for motion estimation uses a spatial-temporal framework which may aid to estimate reasonable motion parameters in a temporally consistent manner from the retrospective measurements (although this aspect is not described in the paper). The method is shown to work on more than one contrast (eg: simulated multi-shot T1 and T2 datasets).
- 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 is fundamentally limited to multi‑shot or segmented MRI acquisitions where motion is approximately rigid and constant within each shot. The standard T1/T2 datasets are not multi-shot acquisitions, where the key assumptions of this paper breaks down. What the paper demonstrates on T1/T2 are simulated cases where a single shot acquisition is split into a multi-shot acquisition retrospectively and does not reflect a standard MRI acquisition. More realistic multi-shot methods would have been multi-shot EPI or non-Cartesian methods (eg: multi-shot spiral, radial), where the feasibility of the method is not demonstrated. The method is stated to be sampling agnostic, however the description makes it clear that the method is sampling aware and is a strict constraint. It is not clear how multiple contrasts or orientations will recast the method as sampling-agnostic.
- 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?
The paper presents a principled method for retrospective motion correction is a well-defined but limited regime of multi-shot acquisitions where the rigid, inter-shot model assumptions hold. The method is tested in those settings using synthetic datasets and analytically simulated motion and is shown to work well. However, neither the standard clinical T1/T2 nor the fastMRI T1/T2 datasets are multi-shot datasets, rather they are single-shot methods where the assumptions do not hold. The method is not validated in a true multi-shot dataset where the acquisition itself was performed in a multi-shot fashion. Moreover the sampling-agnostic claim is not supported in the experiments.
- 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 sincerely thank all reviewers and the meta-reviewer for their positive reviews and constructive feedback, which have been very helpful in improving the quality of our work. Below, we address their main comments as follows:
(1) We agree with R3 that our proposed method is limited to multi-shot or segmented MRI. However, we would like to clarify that the term “multi-shot” in our work includes segmented acquisitions in which multiple k-space lines are acquired within one TR, such as T1-weighted MPRAGE, T2-weighted TSE, and FLAIR sequences. These sequences are widely used in standard clinical scans. In fact, the T1-weighted, T2-weighted, and PD-weighted images from the IXI dataset were acquired in a multi-shot manner, with echo-train-lengths of 208, 16, and 16, respectively. We have added the following text to the revised paper to clarify this point: “Notably, all images were acquired in a multi-shot manner with varying echo-train-lengths.” Additionally, we revised the opening sentence of Section 2.1: “We focus on the retrospective motion correction problem in 2D multi-shot MRI, such as in common sequences like fast spin echo.”
(2) We agree with R3 that the term “sampling-agnostic” may be imprecise, as the initial motion estimation takes sampling patterns as input, making the method “sampling-aware”. Our intended meaning was that SAMoCo is not limited to a single fixed sampling pattern; the same trained network can handle variable shot numbers, ACS lines, and acceleration factors without retraining. In future publications, we will replace “sampling-agnostic” with more precise terminology, such as “sampling-flexible”, to better describe the ability of the framework to generalize across different sampling schemes.
(3) We thank R1 and R2 for their valuable suggestions regarding real-motion validation, generalization to shot numbers beyond the training distribution, non-Cartesian sampling, pathological data, and downstream clinical tasks. We plan to investigate these issues in our future work. In addition, we will follow R1’s suggestion to further analyze the influence of initial motion estimation and joint sensitivity map updating through additional ablation studies.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
The reviewers agree that the paper presents a well-motivated and technically sound hybrid framework that combines learning-based initialization with physics-driven refinement for MRI motion correction. The method is clearly described, shows consistent improvements over strong baselines, and demonstrates promising generalization across datasets and acquisition settings. While concerns remain regarding evaluation scale, realism of motion settings, and some claims (e.g., sampling-agnostic behavior), these are viewed as limitations that can be addressed in future work rather than fundamental flaws. Overall, given the solid methodology, empirical gains, and potential clinical relevance, I recommend acceptance.