List of Papers Browse by Subject Areas Author List
Abstract
Magnetic Resonance Imaging (MRI) is a vital non-invasive imaging modality that is highly sensitive to patient motion, which disrupts k-space consistency and produces structured artifacts in reconstructed images. Most existing deep learning–based motion correction methods operate exclusively in the image domain, implicitly discarding phase information and limiting their ability to correct motion-induced frequency distortions. Although several recent ap-proaches incorporate k-space awareness, they primarily use k-space to detect corrupted lines or impose data-consistency constraints, while motion correc-tion itself remains confined to the image domain. As a result, phase-inconsistent frequency components introduced during acquisition are not ex-plicitly corrected. To address this limitation, we propose KiMoCo-Net, a dual-domain motion correction framework that aligns learning-based correction with the physics of MRI acquisition. KiMoCo-Net consists of two sequential stages: (i) a frequency-domain correction network that operates directly on complex-valued k-space to restore phase and magnitude consistency prior to reconstruction, and (ii) a spatial-domain refinement network that enhances an-atomical structures after inverse Fourier transformation. The model is trained end-to-end using a hybrid objective that jointly enforces k-space fidelity and image-domain structural similarity. Extensive experiments on synthetic and real motion-corrupted MRI data demonstrate that KiMoCo-Net outperforms state-of-the-art image-domain and hybrid methods under severe motion while exhibiting robust performance across motion severities, achieving substantial improvements in structural fidelity and image sharpness (18.2% SSIM and 7.61% PSNR gains). Moreover, KiMoCo-Net maintains low inference time, supporting its practical applicability in clinical settings.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4105_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/marina513/Prior-Mixer-KiMoCoNet
Link to the Dataset(s)
Calgary–Campinas dataset: https://doi.org/10.1016/j.neuroimage.2017.08.021
PMoC3D dataset: https://doi.org/10.48550/arXiv.2506.05975
BibTex
@InProceedings{FekMar_KiMoCoNet_MICCAI2026,
author = { Fekry, Marina Maher AND El-Beltagy, Mohamed A. AND Makary, Meena M. AND Al-masni, Mohammed A.},
title = { { KiMoCo-Net: Dual-Domain Learning for MRI Motion Artifact Correction } },
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
The authors propose KiMoCo-Net, a sequential dual-domain DL framework designed to retrospectively correct motion artifacts in MRI. Unlike most existing methods that operate solely in the image domain, this approach aligns with the physics of MRI acquisition by first correcting phase and magnitude inconsistencies directly in the frequency domain, followed by a secondary refinement step in the image domain to enhance anatomical structures.
- 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.
- Addressing MRI motion artifacts directly in k-space, where the corruption actually occurs during sequential phase-encoding, is logical and superior to purely image-based post-processing.
- The proposed method operates at approximately 0.06 seconds per volume. Compared to the competing SOTA IM-MoCo (which takes ~50 seconds), this is an advantage that supports the authors’ claim of clinical applicability.
- Table 2 provides evidence for the architecture’s design. By comparing the dual-domain approach against single-domain networks and cascaded image-domain networks, the authors successfully prove that the performance gains come from cross-domain complementarity rather than an increase in network depth.
- 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 abstract and conclusion state that KiMoCo-Net “consistently outperforms state-of-the-art [. . . ] methods” across quantitative metrics. However, Table 1 shows that the baseline IM-MoCo achieves a PSNR of 38.80, which is over 6 dB higher than the proposed method’s PSNR of 32.68.
- The SSIM margin between KiMoCo-Net (0.947(0.032)) and IM-MoCo (0.935(0.073)) is relatively small, and their standard deviations overlap. Given the high variance, a statistical significance test (e. g. , wilcoxon signed-rank test) is needed to claim that this improvement is not due to chance.
- The variable I^ is used contradictorily. In Eq. 2, I^ is the output of the image-space network. In Eq. 4, it is suddenly used in the k-space loss function. In Eq. 5, a new variable I^′ appears out of nowhere as the final refined reconstruction. This notation must be cleaned up.
- The SSIM metric is standardly defined for real-valued images. Since the inverse Fourier transform of the k-space output yields complex-valued images (magnitude and phase), the authors must explicitly state how SSIM is computed in Eq. 4 and 5 (e. g. , solely on the magnitude images).
- The BLUR metric is prominently featured in the tables and text, yet it is never mathematically defined or referenced in the methodology section.
- 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.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the dual-domain concept is sound and the inference speed is fast, the manuscript in its current state suffers from flaws in scientific reporting, e.g., making claims of outperforming all baselines while ignoring a 6 dB PSNR deficit to a directly compared SOTA method. Furthermore, the mathematical formulation is sloppy, and essential definitions (like the BLUR metric and the handling of complex SSIM) are missing.
- 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.
I thank the authors for their transparent and constructive rebuttal. The stratification by motion severity clarifies the PSNR gap and provides a more honest picture of the method’s strengths. Fixing the mathematical notation and metric definitions addresses my remaining concerns. Given the method’s fast inference speed and the now properly contextualized robustness under severe motion, I am happy to raise my score.
Review #2
- Please describe the contribution of the paper
The paper introduces KiMoCo-Net, a deep learning framework for MRI motion artifact correction. Unlike prior approaches that operate exclusively in the image domain, either with or without implicit k-space guidance, KiMoCo-Net adopts a two-stage strategy: initial correction is performed directly in k-space, followed by refinement in the image domain. This represents a fundamentally different paradigm, explicitly targeting phase and frequency-domain inconsistencies introduced during acquisition, which are only indirectly addressed in previous methods.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper introduces a methodologically distinct approach compared to prior work, operating directly in k-space with subsequent refinement in the image space. This contrasts with existing methods that either operate exclusively in the image space or rely on implicit k-space supervision while still performing corrections in the image domain.
2.Table 1 compares KiMoCo-Net against 5 well-established baselines, demonstrating improved SSIM, PSNR, and BLUR metrics on a synthetic test cohort of 1200 patients with varying simulated motion artifacts, including rotations, translations, and the number of motion events.
3.Table 2 provides compelling ablation studies on simulated data that support the paper’s hypothesis: corrections performed solely in k-space or solely in the image domain are inferior to a two-stage approach, where k-space correction is followed by image-space refinement, thereby justifying the proposed design.
- 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.Quality of simulations (Eq.1): The main premise of the paper is that “By aligning the correction process with the physical origin of motion artifacts, KiMoCo-Net enables more faithful signal recovery and improved image fidelity compared to image-domain-only or k-space-guided methods.” That is, the authors assume that motion corrupts the signal directly in k-space during acquisition. However, Eq. 1 models motion by applying a motion operator M_n to a clean image I_GT, followed by a Fourier Transform. This effectively approximates the process as image -> motion -> FFT rather than the physically grounded sequence of motion during acquisition -> corrupted signal -> corrupted k-space. This discrepancy raises questions about how well the simulation reflects true k-space corruption. It would be valuable for the authors to comment on the validity of this approximation and its implications for faithfully modeling real-world motion artifacts.
2.Quality of simulations and evaluation metrics (SSIM/PSNR as a function of motion severity): In Figure 2, the SSIM values for the corrupted images (red / Artifact) are not monotonic with respect to motion severity. That is, the median SSIM is approximately 0.65 for mild artifacts, increases to 0.90 for moderate corruption, and then decreases to 0.78 for severe corruption. This behavior is counterintuitive, as one would expect SSIM to degrade monotonically with increasing corruption. Clarification on this trend would help assess the validity of the simulation process defined in Eq. 1.In contrast, PSNR exhibits a more expected trend, with median values of roughly 30, 30, and 24 for mild, moderate, and severe corruption, respectively. While closer to monotonic, the relatively limited separation between mild and moderate cases raises concerns about whether PSNR provides sufficient sensitivity to distinguish different levels of motion degradation. It may be beneficial to consider additional metrics with greater dynamic range across severity levels to more effectively characterize both the simulated artifacts and the model’s performance.
3.Real-world deployment and relevance: The paper primarily evaluates model performance on a test set of 1200 simulated patients generated via Eq. 1, with limited additional validation on real-world motion-corrupted data. According to Reference 14, the PMoC3D dataset consists of only eight patients, each with one clean MRI scan, three motion-corrupted acquisitions, and corresponding raw k-space data, making it consistent with the paper’s underlying assumptions. However, the very small sample size substantially limits statistical power, and the evaluation on this dataset is restricted to qualitative results (Figure 4), with no accompanying quantitative metrics. Finally, the proposed approach requires access to both raw k-space data and reconstructed images, effectively relying on paired data that are often difficult to acquire in practice. This requirement may pose a significant barrier to large-scale data collection for training and evaluation, especially when compared to image-only methods. As a result, the approach may remain dependent on synthetic data, which, given the concerns raised in Points 1 and 2, could further limit its real-world applicability.
4.Choice of Generative baseline: As illustrated in Figure 1, the problem can be naturally formulated as a supervised translation task, given the availability of paired corrupted and clean k-space data. In such settings, paired image-to-image generative translation frameworks, ex. pix2pix, are typically more appropriate. However, the authors instead adopt CycleGAN as a generative baseline, which is designed for unpaired data. This choice is not clearly justified and may be suboptimal, as CycleGAN does not fully leverage the paired supervision available in this setting. Clarification on this design decision, and potentially comparison to paired approaches, would strengthen the evaluation.
- 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.
(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 provides a clear and comprehensive overview of the motion correction landscape and introduces a novel approach that performs correction directly in k-space, rather than in the image domain with implicit k-space regularization. Through comparisons with state-of-the-art methods and supporting ablation studies, the authors demonstrate that a two-stage strategy, k-space correction followed by image-space refinement, can be beneficial. This constitutes the paper’s primary strength, offering a conceptually meaningful contribution that advances prior work.
However, several limitations remain regarding underlying assumptions, evaluation, and real-world applicability. Notably, the formulation in Eq. 1 appears inconsistent with the paper’s central claim of addressing the “physical origin of motion artifacts,” as it simulates motion in the image domain prior to transformation into k-space. Additionally, the non-monotonic SSIM trends observed in Figure 2 raise concerns about the validity of the simulation framework and its suitability for evaluating performance across varying levels of corruption. Finally, the requirement for paired clean and corrupted k-space data presents challenges for real-world deployment and scalability, particularly given the limited size of available datasets (e.g., N=8 for PMoC3D).
If these concerns, particularly those related to simulation fidelity, evaluation robustness, and data requirements, are adequately addressed, the proposed approach has the potential to represent a meaningful methodological advancement over existing techniques.
- Reviewer confidence
Somewhat confident (2)
- [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.
Comments on simulation fidelity, visualization errors, real-data deployment, and the selection of generative baselines were adequately addressed.
Review #3
- Please describe the contribution of the paper
This paper proposes KiMoCo-Net, a dual-domain deep learning framework for MRI motion artifact correction. The method uses a k-space correction network operating on the real and imaginary components of complex-valued corrupted k-space, followed by inverse Fourier transformation and an image-domain refinement network. The method is that explicit k-space correction enables better restoration of phase and frequency consistency than image-domain-only or k-space-guided methods, leading to improved robustness, especially under severe motion.
- 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.
This work has a clean, modular architectural design. The two-stage pipeline is easy to understand, and the diagram clearly communicates how corrupted k-space is first corrected and then transformed for image refinement. The method is tested on simulated motion corruption at multiple severity levels and is also qualitatively shown on real motion-corrupted MRI. The comparison includes several relevant baselines, and the ablation study is useful.
- 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.
Several technical details are underspecified, which makes it difficult to fully assess reproducibility. For example, it is not entirely clear whether training and inference are performed slice-wise or volumetrically, how the “complex-valued” k-space network is implemented, or whether the model uses true complex convolutions or simply concatenates real and imaginary channels.
- 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?
This paper has a clear core idea and a technically reasonable formulation. The method is aligned with MRI acquisition physics, and the dual-domain design is both intuitive and potentially useful in practice. The reported robustness under severe motion and favorable runtime also makes the work practically interesting.
However, some implementation details remain unclear, and the empirical case, while promising, is not yet fully airtight. The real-data evaluation remains limited, and the fairness and exact setup of the baseline comparisons require clearer documentation.
- 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
We thank the reviewers for their feedback. R1.1: PSNR Superiority We agree that the manuscript wording should be more precise. While IM-MoCo achieves a higher average PSNR, this is mainly driven by mild motion cases (51.13±8.66 dB vs. 35.29±3.30 dB for KiMoCo-Net). Under severe motion, KiMoCo-Net performs better (30.25±1.87 vs. 28.82±4.61). IM-MoCo shows higher variance across severities (11.79 vs. 3.28). We will clarify this interpretation and avoid implying uniform superiority across all metrics.
R1.2: Statistical Significance We performed Wilcoxon signed-rank tests. Results show severity dependence: KiMoCo-Net significantly outperforms IM-MoCo in SSIM under severe motion (0.914±0.027 vs. 0.869±0.068; p=5.7×10⁻⁴⁰), while IM-MoCo excels in mild motion (0.990 vs. 0.978; p=2.15×10⁻³⁷). Consequently, the overall SSIM difference is not statistically significant (p=0.06). Thus, KiMoCo-Net exhibits greater robustness under severe motion, with similar trends observed for PSNR (p<0.05). We will clarify this interpretation.
R1.3: Notation, SSIM, and BLUR We acknowledge the notation inconsistency: I^ should be I^′ in Eq. (2), where I^ denotes the intermediate reconstruction after inverse FFT and I^′ the final image-domain refinement. We will correct Eq. (2) and clarify the notation in Fig. 1.SSIM is computed on magnitude images after applying the absolute operation. We will also explicitly define the BLUR metric following Crété-Roffet et al. (2007).
R2.1: Simulation Realism Eq. (1) follows established MRI motion simulation frameworks (Batchelor et al., MRM 2005). Motion occurs in the image/object domain, while k-space is acquired sequentially during sampling. Thus, each motion state M_n alters the object prior to Fourier encoding, causing different phase-encoding lines to be acquired from different spatial states. We agree that Eq. (1) approximates continuous motion during acquisition; however, it is widely adopted in retrospective MRI motion simulation and captures the key mechanism underlying motion-induced k-space inconsistencies. Qualitative generalization on PMoC3D supports this formulation.
R2.2: Non-monotonic Trend We thank the reviewer for identifying this issue. The boxplots contained visualization errors. Correct artifact scores are: mild (SSIM=0.90, PSNR=30.75), moderate (0.758, 23.86), severe (0.63, 20.59). U-Net scores: mild (SSIM=0.96, PSNR=33.32), moderate (0.911, 29.81), severe (0.84, 27.08). We will correct the corresponding plots.
R2.3, R3: Real Data Evaluation We agree that PMoC3D is limited in size. Nevertheless, it was included to assess qualitative generalization beyond synthetic motion simulation. Quantitative evaluation remains challenging because motion-free ground truth is unavailable, the provided “clean” scans still contain residual motion, and existing no-reference metrics are unreliable for separating true anatomy from motion-induced artifacts. Despite these limitations, the proposed method consistently showed visually improved artifact suppression and clearer anatomical preservation across multiple patients and slices compared to competing methods, suggesting good generalization to real motion-corrupted MRI. Additional examples will be added to Fig. 4. R2.4) Pix2Pix Results We selected CycleGAN to represent GAN-based image-domain correction methods due to its widespread use in prior MRI motion correction literature. The Pix2Pix results will be added to Table 1: SSIM (0.821, 0.722, 0.604, avg 0.715) and PSNR (25.4, 22.9, 20.6, avg 22.9). R3: Implementation Details/Reproducibility We will expand implementation details in the revised manuscript, and the code will be released upon acceptance. Training/inference is performed slice-wise. The k-space network processes concatenated real and imaginary channels using standard real-valued convolutions rather than complex convolutions. All baselines were trained under identical preprocessing, splits, optimizer settings, and epochs to ensure fair comparison.
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.
The paper presents a potentially interesting dual-domain framework for MRI motion correction that aligns with acquisition physics and demonstrates promising efficiency and robustness. However, the reviewers raise several substantive concerns regarding the validity of the simulation setup, clarity of the formulation, and inconsistencies in performance claims (e.g., PSNR comparison with prior work), as well as limited real-data evaluation. Given these mixed but potentially addressable issues, I invite the authors to provide a rebuttal clarifying these points and strengthening the empirical and methodological justification.
- 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 responded to most of the reviewers’ comments. However, I think there are additional limitations the authors should have discussed. The evaluation on real data is missing. what if no k-space data is avavilable for motion correction in real data? etc…
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.
All reviewers suggest accept with a strong clarification of the previous concerns. The authors successfully resolved initial concerns regarding simulation fidelity and cross-domain generalization
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 paper proposes KiMoCo-Net, a dual-domain motion correction framework that aligns learning-based correction with the physics of MRI acquisition. KiMoCo-Net consists of a frequency-domain correction network that operates directly on complex-valued k-space to restore phase and magnitude consistency, and a spatial-domain refinement network that enhances anatomical structures after inverse Fourier transformation. The reviewers all endorsed the strength of the paper including incorporating the MRI physics in the frequency domain. The concerns regarding claims of performance gain, simulation setup, significance tests, baseline selections, and real-world evaluations are well addressed after rebuttal. Thus, I am giving an acceptance recommendation.
