List of Papers Browse by Subject Areas Author List
Abstract
Motion artifacts degrade MR image quality affecting clinical diagnoses. Although deep learning-based motion artifact correction (MAC) methods show promise, they are limited by the lack of real paired motion-corrupted and motion-free images. We propose a novel frequency-assisted artifact disentanglement learning framework for MAC of MR images. Our approach integrates a frequency-decomposed motion correction network (FDMC-Net) for content-artifact disentanglement over the real unpaired data, coupled with confidence-guided knowledge distillation using simulated paired data. Specifically, considering that motion artifacts are primarily caused by high-frequency k-space misalignment, FDMC-Net decomposes motion-corrupted MR images into low-frequency and high-frequency components and then employs dedicated encoders to disentangle content and artifact features. FDMC-Net is trained by unsupervised cycle-consistent adversarial loss over realistic unpaired data, and confidence-guided knowledge distillation loss by distilling a teacher model trained on simulated paired data. Experiments demonstrate its state-of-the-art performance, with ablation studies confirming the effectiveness of frequency-assisted disentanglement and confidence-guided distillation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0954_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{WanJia_MRI_MICCAI2025,
author = { Wang, Jiazhen and Yang, Heran and Yang, Yizhe and Sun, Jian},
title = { { MRI Motion Artifact Correction via Frequency-Assisted Artifact Disentanglement and Confidence-Guided Knowledge Distillation } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15972},
month = {September},
page = {392 -- 401}
}
Reviews
Review #1
- Please describe the contribution of the paper
The manuscript introduces a Frequency-Assisted Artifact Disentanglement framework for motion artifact correction in MRI. The proposed method integrates a Frequency-Decomposed Motion Correction Network (FDMC-Net), which separates high- and low-frequency information, with a Confidence-Guided Knowledge Distillation (CGKD) module. The framework aims to address the challenge of lacking real paired motion-corrupted and motion-free MRI data by leveraging simulated artifacts for guidance while applying the correction to unpaired real-world data.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The paper tackles an important and clinically relevant problem: correcting motion artifacts in MRI where clean ground truth data is often unavailable.
- The idea of combining real unpaired data with simulated data in a semi-supervised manner through knowledge distillation is well-motivated and promising.
- 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.
- Figure 1 presents Imf and Imc as inputs, but the proposed method claims to work on unpaired real data. Clarification is required on how the model handles these inputs in practice under the unpaired setting.
- Section 2.1 lacks clarity in explaining how the ‘M’ block separates content (c) from artifact (a). A more intuitive explanation would enhance understanding.
- The paper does not reference or compare its method to a relevant recent work on disentangled motion artifact correction presented at MICCAI 2024 (https://doi.org/10.1007/978-3-031-72114-4_21). A comparison or at least a discussion is needed to highlight the novelty and difference of the proposed approach.
- It is unclear whether the same patient data is used in both the simulated and real inputs for Figure 1 (a) and (b). If not, it raises concerns about the relevance and transferability of the distilled features. If yes, the claim of working in an unsupervised setting may not be fully accurate and should be discussed.
- It is not specified whether the simulated motion data (SM1 and SM2) are derived from an existing dataset (e.g., MR-ART) or self-generated. If the latter, detailed descriptions of the simulation process and motion parameters are necessary.
- Based on Table 1, the SSIM values for corrupted inputs (SM1: 0.8523, SM2: 0.8795, RM1: 0.9185, RM2: 0.8756) suggest relatively mild/moderate degradation. For a robust evaluation, more severe motion artifacts should be simulated and tested. Without this, it is hard to judge clinical applicability.
- Although the method emphasizes the role of frequency decomposition, there is no ablation study assessing the contribution of the low- and high-frequency components.
- Table 1 suggests that corrupted in SM2 has better SSIM than SM1, which contradicts the textual definition (SM2 = heavy motion). The authors should check and possibly switch the columns in Table 1.
- Optional: The inclusion of an additional evaluation metric such as the Blurriness Index would offer complementary insight into image quality, particularly relevant in motion correction tasks.
- 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.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — could be 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 solution to the problem of MRI motion artifact correction using unpaired data, incorporating both frequency decomposition and knowledge distillation in a well-structured framework. However, several important technical clarifications and architectural evaluations (e.g., ablation studies) are needed to fully validate the method’s novelty and effectiveness. Addressing these concerns will greatly enhance the scientific merit and impact of the paper.
- 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 clarified and addressed my comments.
Review #2
- Please describe the contribution of the paper
This paper presents a novel frequency-assisted artifact disentanglement framework for correcting motion artifacts in MR images. The proposed method incorporates a Frequency-Decomposed Motion Correction Network (FDMC-Net), which decomposes motion-corrupted MR images into high-frequency (HF) and low-frequency (LF) components to effectively suppress motion artifacts while preserving structural fidelity. The FDMC-Net is trained on real, unpaired data and guided by a teacher model that trained on simulated pseudo-pairs, using a confidence-guided knowledge distillation (CGKD) strategy to better retain 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.
(1) The task addressed in this paper has valuable practical applications. (2) The idea of using a teacher model pre-trained on simulated paired data to guide the motion correction network is well-motivated and effectively circumvents the challenge of obtaining real paired data. It also mitigates the risk of anatomical distortion or the generation of pseudo-tissues, which is common when training on unpaired data. (3) The decomposition of motion-corrupted MR images into high-frequency (HF) and low-frequency (LF) components is a good trick for disentangling motion artifacts from contextual information. (4) The methodology and experimental setup are clearly described and presented in sufficient detail.
- 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 experimental evaluation in this paper is conducted using only a single dataset (MR-ART), in which all subjects’ HeadMotion1 and HeadMotion2 scans involve fixed-pattern motion—specifically, participants were instructed to nod their heads (tilting down and then up along the sagittal plane) once each time the word “MOVE” appeared on the screen [1]. This controlled and repetitive motion pattern limits the ability to demonstrate the proposed method’s effectiveness in handling complex and varied real-world motion artifacts. [1] Nárai, Á., Hermann, P., Auer, T., Kemenczky, P., Szalma, J., Homolya, I., Somogyi, E., Vakli, P., Weiss, B., Vidnyánszky, Z.: Movement-related artefacts (mr-art) dataset of matched motion-corrupted and clean structural mri brain scans. Scientific data 9(1), 630 (2022)
- 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.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
(1) The inclusion of validation experiments on motion-corrupted data from other datasets (using a motion correction network trained on MR-ART) helps to demonstrate the generalization capability of the proposed method. (2) The visualization results currently include only axial slices; it would be beneficial to also provide results in the coronal and sagittal planes for a more comprehensive evaluation.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed method is well-motivated, intuitive, and inspiring. Experimental results show that it outperforms other unpaired image translation methods. From an MR imaging perspective, both the motivation and experimental design are rigorous. For example, the use of a forward model for motion simulation and the incorporation of prior knowledge that motion artifacts predominantly originate from high-frequency k-space misalignments. That said, the authors are encouraged to address the aforementioned limitations to further strengthen the work.
- 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 primary contribution of this manuscript is the introduction of a novel method named FDMC-Net (Frequency-Assisted Artifact Disentanglement with Confidence-Guided Knowledge Distillation), designed specifically to address motion artifact correction in magnetic resonance imaging (MRI). FDMC-Net uniquely decomposes MRI images into low-frequency (LF) and high-frequency (HF) components to separately extract anatomical content and motion-related artifacts. Additionally, the authors integrate unsupervised cycle-consistent adversarial learning using unpaired real data, along with a confidence-guided knowledge distillation strategy employing synthetic paired data. This combination effectively addresses the scarcity of paired artifact-free reference data in clinical MRI scenarios.
- 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.
-
Integration of Frequency-Aware Disentanglement for Motion Artifact Suppression The proposed FDMC-Net employs a frequency decomposition module to explicitly separate low-frequency (LF) and high-frequency (HF) components of motion-corrupted MR images. By assigning content and artifact encoders to each frequency band, the method facilitates precise disentanglement of anatomical structures and motion artifacts. This formulation is grounded in domain knowledge that motion artifacts predominantly distort high-frequency k-space regions and represents a structured advancement over existing one-stream CycleGAN variants.
-
Confidence-Guided Knowledge Distillation with Pixel-wise Reliability Weighting The distillation framework introduces a stochastic mixture-of-experts teacher model that produces multiple pseudo-labels per input. The pixel-wise variance across these outputs is used to generate a confidence map, which adaptively weights the distillation loss. This mechanism effectively down-weights unreliable pseudo-labels and emphasizes structurally consistent regions, improving robustness against the domain gap between synthetic and real MRI artifacts.
-
Consistent Quantitative Superiority and Well-Designed Ablation Analyses The method demonstrates consistent improvements in PSNR and SSIM metrics across simulated and real motion conditions when compared with state-of-the-art baselines (e.g., CycleGAN, DUNCAN, DCGAN-MS). Comprehensive ablation studies further validate the individual contributions of the frequency decomposition and the confidence-guided distillation. The empirical design is systematic, and each component’s effectiveness is substantiated by interpretable results.
-
- 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.
- Over-simplified Assumption That Motion Artifacts Are Confined to High Frequencies
The method presumes that motion artifacts are predominantly localized in the high-frequency domain of the MR image, while low-frequency components remain largely unaffected and represent clean anatomical structures. While this assumption may hold for certain rigid body motions, it fails to account for more complex or large-scale deformations such as bulk motion, spin history effects, or motion during low-frequency k-space sampling—which can induce significant distortions in the low-frequency components as well.
- Recommendation: The authors are encouraged to conduct a frequency-domain error analysis (e.g., power spectrum difference between clean and corrupted images) to empirically support the HF artifact assumption. Additional experiments on samples with low-frequency-dominant artifacts or introducing adaptive or multi-band frequency decomposition (e.g., wavelet-based) would strengthen the argument and test the robustness of their architecture under broader motion scenarios.
- No Direct Evidence Supporting the Claimed Disentanglement of Content and Artifact Features
While the architecture is explicitly designed to disentangle content and artifact representations through separate encoders for each frequency band, the manuscript does not provide any qualitative visualization or interpretability-driven validation of this disentanglement. It remains unclear whether the model has indeed learned to isolate anatomical content from motion-induced distortions, as the disentangled features are never visualized or examined independently.
- Recommendation: To substantiate this central claim, the authors should include visualizations of the encoded content and artifact features (e.g., by decoding each component individually and inspecting the resulting images). Additionally, feature map overlays, PCA or t-SNE plots showing cluster separation between content and artifact embeddings, or attention heatmaps indicating what each encoder focuses on would lend credibility to the disentanglement claim. Without such analysis, the network’s internal behavior remains opaque, and the term “disentanglement” lacks empirical support.
- Lack of Verification of the Teacher Model’s Pseudo-label Quality
The confidence-guided distillation loss relies entirely on pixel-wise variance across T stochastic forward passes of the teacher model as a surrogate for pseudo-label reliability. However, low variance does not necessarily indicate correctness; confidently wrong pseudo-labels may still dominate training, especially in anatomically complex or low-contrast regions.
- Recommendation: The authors should evaluate and report the absolute accuracy (e.g., PSNR, SSIM) of the teacher model on the synthetic validation set, and visually compare teacher predictions with ground-truth clean images. Furthermore, implementing pseudo-label filtering techniques such as entropy-based thresholding or variance + error margin estimation could enhance distillation robustness. These enhancements would prevent over-reliance on potentially flawed supervision.
- Limited Anatomical and Modal Generalizability
The method is evaluated exclusively on T1-weighted brain MRI from the MR-ART dataset. This narrow scope raises concerns about its applicability to broader clinical contexts, where motion artifact characteristics vary significantly across imaging protocols and anatomical regions (e.g., abdominal, spine, cardiac MRI). As such, the paper’s claim to clinical feasibility is weakened by the lack of external validation.
- Recommendation: To support broader applicability, the authors should perform experiments on additional MRI contrasts (e.g., T2-weighted) and anatomical domains (e.g., abdomen, heart). In the absence of real data, simulated motion artifacts on publicly available datasets (e.g., IXI, HCP, fastMRI) could be employed as a reasonable proxy.
- Over-simplified Assumption That Motion Artifacts Are Confined to High Frequencies
The method presumes that motion artifacts are predominantly localized in the high-frequency domain of the MR image, while low-frequency components remain largely unaffected and represent clean anatomical structures. While this assumption may hold for certain rigid body motions, it fails to account for more complex or large-scale deformations such as bulk motion, spin history effects, or motion during low-frequency k-space sampling—which can induce significant distortions in the low-frequency components as well.
- 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.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I am assigning a score of 4 (Weak Accept) to this manuscript. The paper presents an innovative and technically sound approach for MRI motion artifact correction using a combination of frequency-aware artifact disentanglement and confidence-guided knowledge distillation. While the proposed methodology is well-motivated and the empirical results are strong, several critical limitations in validation, interpretability, and generalizability prevent a higher recommendation at this stage. Below, I detail the primary factors influencing my decision.
[Reasons in Favor of Acceptance]
-
Conceptual Novelty and Architectural Design The introduction of a frequency-decomposed motion correction network (FDMC-Net) that explicitly disentangles content and artifact features in both low- and high-frequency domains is a compelling innovation. The dual-encoder structure for content and artifact separation represents a clear architectural advance over prior CycleGAN-style approaches. Furthermore, integrating a confidence-aware distillation mechanism from a teacher model via pixel-wise variance maps adds an additional layer of robustness, particularly in handling domain gaps between synthetic and real motion artifacts.
-
Strong Quantitative Results and Ablation Studies The method achieves consistently superior performance across multiple motion scenarios (simulated and real) on the MR-ART dataset, outperforming strong baselines such as CycleGAN, DUNCAN, and DCGAN-MS. The authors further conduct systematic ablations that demonstrate the importance of both the frequency decomposition and the confidence-guided distillation components, reinforcing the validity of their design choices.
-
Relevance and Practicality in Clinical MRI The approach directly addresses a well-known and challenging problem in clinical MRI—the lack of paired motion-free/motion-corrupted data—and presents a strategy that leverages both unpaired real-world data and simulated pairs. The emphasis on unpaired training and robustness to motion type reflects practical applicability and aligns with the MICCAI community’s interest in clinically translatable solutions.
[Reasons Preventing a Higher Recommendation]
-
Over-reliance on Frequency Assumptions The entire framework assumes that motion artifacts are concentrated in the high-frequency domain. While grounded in domain knowledge, this assumption is not empirically tested in the paper, and may not generalize to all types of motion. For example, some artifacts—such as large-scale drift or spin history effects—can degrade low-frequency signal components, which are currently treated as purely anatomical in the model.
-
Disentanglement Not Empirically Verified Although the model is architecturally designed to separate content and artifact features, no qualitative or quantitative evidence is provided to confirm that this separation is actually achieved. Without visualizations or interpretability analysis (e.g., reconstructed images from content/artifact alone, t-SNE feature clustering), the core claim of disentanglement remains speculative.
-
Unverified Reliability of Pseudo-labels from the Teacher Model The confidence-guided knowledge distillation relies solely on the variance of multiple pseudo-labels for reliability estimation but does not assess the accuracy or potential bias of the teacher model itself. This may introduce a risk of overfitting to confidently incorrect labels, especially in challenging anatomical areas.
-
Limited Validation Beyond a Single Dataset and Contrast The experimental validation is restricted to T1-weighted brain MRI from the MR-ART dataset. There is no evidence of generalizability to other imaging protocols (e.g., T2, FLAIR) or anatomical regions (e.g., abdomen, cardiac), which limits the clinical scope of the claims. Given the model’s complexity and intended real-world use, broader evaluation is essential.
-
- 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 thank all the reviewers for their comments and will revise accordingly. [R1-Q1,Q4] Upaired and Unsupervised Setting. In Fig. 1a, motion-free image Imf and motion-corrupted image Imc are from different subjects. Using the same subject’s data as student model in Fig.1a, the teacher model in Fig.1b produces pseudo-labels and confidences for student training. Our full model is trained on unpaired Imf and Imc without using paired motion-free image of Imc, ensuring an unpaired and unsupervised setting. [R1-Q2,Q7,R3-Q2] Disentanglement Details and Validation. The frequency decomposition module (FDM) employs dynamically gated learnable filters to extract low-frequency (LF) and high-frequency (HF) components using Eqn. (1). These components are then fed into LF and HF content encoders and LF and HF artifact encoders to generate the LF & HF content features and LF & HF artifact features in Eqn. (2). T-SNE visualization in Fig. 3 shows well separation between content and artifact features. Table 2 shows
w/o CGKD’’ (retaining FDM and removing CGKD) improves PSNR by 1.25 dB over
baseline’’ (removing FDM and CGKD), showing the effectiveness of FDM. Please refer to [R3-Q1] for more discussion. [R1-Q3] Comparison with Multi-Net. Multi-Net is a multi-task framework for motion artifact correction (MAC) and tissue segmentation. For MAC, it uses dual-domain encoders to separate content and artifact features. In contrast, our method disentangles content and artifacts in both LF and HF components, enabling artifact separation and suppression while preserving content features (see Table 2,w/o CGKD’’ vs.
baseline’’). We also introduce the confidence-guided distillation for synthetic-to-real knowledge transfer to further improve accuracy (see Table 2,Ours’’ vs.
w/o CGKD’’). We will cite Multi-Net and include the discussion. [R1-Q5,Q8] Simulation Clarification. SM1 and SM2 are directly from the MR-ART. SM1(pitch15dur2p5nnods10’’) simulates 10 nods with fixed 15° pitch rotation. SM2 (
rot0to15nnods5’’) simulates 5 nods with random 0-15° rotation across the pitch, yaw, and roll axis. We will clarify the simulation settings of SM1 and SM2 in the paper. [R1-Q6,R2-Q1] Evaluation under Severe Motion. While Table 1 shows average results, each dataset includes severely degraded cases due to random simulation parameters and severe subject motion in real scans. On selected 5 subjects (avg. SSIM=0.679) from RM2 with severe motion, our method outperforms the second-best method DCGAN-MS [4] by 0.076. As new experiments are not allowed, more evaluation will be conducted in future. [R3-Q1] Frequency Analysis. We validate our frequency-based design via frequency error analysis, which reveals that artifacts are mainly in HF components, and with noticeable LF errors under severe motion. Our method performs content-artifact disentanglement within both LF and HF components, by extracting LF & HF content and LF & HF artifact features using specialized encoders in Eqn.(2). On 10 subjects selected from SM2 with strongest LF-dominant artifacts, our method outperforms the second-best method DCGAN-MS [4] by 0.088 in SSIM. [R3-Q3] Pseudo-label Quality and Reliability. The trained teacher model underperforms our full model by >1dB. Based on the pseudo-labels, we introduce a confidence map to focus on reliable regions and incorporate a cycle-consistency adversarial loss as an unsupervised constraint. New experiments are not allowed, the suggestions on pseudo-label filtering strategies are appreciated and will be explored in future. [R3-Q4] Modality and Anatomy Generalization. We have validated our method on two simulated and two real T1-weighted brain datasets. Since the core components are not restricted to specific modality or anatomy, our method can be directly applied to other MRI modalities or anatomical regions by retraining or training over mixed datasets. Due to rebuttal constraints, we will work on other anatomies and modalities in future.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
All reviewers agree that the method is well motivated and intuitive, and the validation is sound and shows that the method outperforms baselines. While some limitations are noted (assumption that artefacts are restricted to low frequencies, lack of ablation study to evaluate the impact of the many losses and components of the method, evaluation on real but controlled motion trajectories, etc), the paper is overall of high quality and clearly passes the bar for presentation at MICCAI.