acceptodds
Withdrawn from ICLR 2027

MARMI: Marginalizing Anatomical Uncertainty for MRI Motion Inference

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Patient motion in multi-shot magnetic resonance imaging (MRI) introduces k-space inconsistencies that depend jointly on the motion trajectory and the underlying anatomy, making the two difficult to disentangle from the acquired measurements. A motion-free auto-calibration signal (ACS) anchors the reference pose but only partially constrains the corresponding anatomy, leaving motion estimation intrinsically ambiguous. We introduce MARMI (Marginalized Anatomical References for Motion Inference), a framework that models ACS-conditioned anatomical uncertainty and marginalizes it during amortized shot-wise motion inference on a compact central k-space representation. MARMI consists of two complementary components. Specifically, Conditional Anatomical Residual Sampling (CARS) generates plausible motion-free references by modeling the anatomical content not determined by the ACS, providing a direct physical interpretation of the unobserved frequencies. Reference-Marginalized Motion Inference (RAMI) then evaluates these references through a motion-aware MRI forward model and aggregates their physical evidence for motion estimation. On fastMRI FLAIR, marginalizing multiple ACS-conditioned references reduces normalized motion error by 10–13% relative to a single sampled reference. MARMI also outperforms JSMoCo and MoPED in motion-corrected reconstruction across the tested ACS and acceleration settings, and extends to 0.3 T M4Raw and an adapted in-house 0.55 T cohort. Together, these results show that modeling the partially observed reference as a latent distribution provides a principled strategy for disentangling motion from anatomy.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
MRI motion correction, motion estimation, physics-informed inference, variational inference, inverse problems

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_WiSYOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors propose MARMI, a MRI rigid motion correction framework for marginalizing over band-limited clean reference images, with the main argument being that existing motion correction approaches ignore uncertainty and focus on point estimates of clean reference images. This framework consists of two steps: (1) CARS, which generates plausible band-limited samples conditioned on the fully sampled ACS region using a conditional diffusion model, and (2) RAMI which uses a transformer to predict the motion state from each of the shots and band-limited samples. Results on simulated motion indicate that MARMI outperforms two baselines, MoPED and JSMoCo.

Critical strengths

Motion correction in MRI is an important problem and having proper uncertainty estimates of the underlying image (even if band-limited) can help improve trust and reduce motion artifacts in real-world scenarios.

Critical weaknesses

The authors claim Bayesian marginalization of a low-res clean image is the main contribution. However, they opt to train a transformer to estimate reference weights and motion parameters instead of using the analytically available likelihoods for optimization. There is no demonstrated benefit for this approach which is less interpretable, when the motion is so low-dimensional under the rigid parameterization. Additionally, the reference weights are not explicitly used to estimate the motion in the transformer, but rather output in parallel, which is concerning.

The claim that existing methods do not account for uncertainty in the reference image is not true; for example, diffusion-based rigid motion correction [1] the rigid motion parameters θ\theta and clean image xx are jointly sampled using diffusion-based posterior sampling, with each diffusion posterior sample representing a new draw of the clean image posterior.

Finally, the writing is difficult to follow and often overcomplicates things. Section 2.3 on CARS spends a full paragraph defining a complex linear map and describing properties of linear operators, just to lead to the residual of the band-limited image. Section 2.4 and 2.5 are very difficult to follow, with many variables introduced.

[1] Levac et al. Accelerated motion correction with deep generative diffusion models, MRM 2024

Questions for authors

How does using the analytic marginalized likelihood for the RAMI step compare to the transformer?

Other methods that also jointly optimize the image and motion estimate continuous update the estimated image; however, MARMI samples KK clean image candidates at the start and fixes them - what if none of these KK are close to correct (maybe due to measurement noise, smaller ACS, etc.)? Does it impact the performance of RAMI?

Reviewer_fCx1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper considers an important problem in MRI reconstruction. Due to the long acquisition time of MRI, motion artifacts can be hard to avoid, which makes this problem important. The paper proposes MARMI to address this problem, and the experiments show that the proposed method performs better than the baseline methods. However, the experimental setup may be too simple and unrealistic, and it may not convincingly reflect the effectiveness of the proposed method.

Critical strengths

S1 The paper conducts clear experiments in a 2D setup, and the results show that the proposed method provides better reconstruction quality than the baseline methods.

S2 The paper is well presented, and the ablation study is clear and informative.

Critical weaknesses

W1 The main weakness of this paper is that the experimental setup is too simple to reflect the paper’s contribution. For example, the paper only considers the 2D MRI dataset, and the motion is 2D rigid motion with limited rotation and translation. Therefore, the problem itself may not reflect real-world challenges. Specifically:

  • Real-world motion is almost always 3D, so a 3D MRI dataset may be more appropriate for evaluation.
  • The paper only considers simulated motion, which may underestimate the difficulty of the problem. It would also be good to consider a dataset with real-world motion, or to sample real-world motion for a volume, to validate the effectiveness of the proposed method.
  • Although the inter-shot motion model is reasonable, considering intra-shot motion would make the method more practical.

W2 The paper could include more baselines, including recent state-of-the-art methods as well as classical methods such as alternating optimization methods [1]. Classical methods can usually estimate motion very well, so such a comparison would help demonstrate that the proposed method outperforms classical optimization-based methods.

[1] Cordero-Grande, Lucilio, et al. "Sensitivity encoding for aligned multishot magnetic resonance reconstruction." IEEE Transactions on Computational Imaging 2.3 (2016): 266-280.

Questions for authors

Comparison of estimated motions The differences between methods could come from either the motion estimation or the reconstruction. Is the proposed method also better at motion estimation? It would be helpful to provide a comparison of the estimated motions across different methods.

JSMoCo implementation details The paper mentions that JSMoCo uses a k-space phase-ramp motion operator. Could the authors clarify how this motion operator is implemented? If only the phase ramp is used, the method only accounts for translation and not rotation, which may underestimate its performance.

Reviewer_gRQ7OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses retrospective motion correction in multi-shot Cartesian brain MRI, where the head is assumed to undergo rigid, in-plane motion that is constant within each shot. The method uses the central auto-calibration signal (ACS), assumed to be acquired without motion, as a reference pose and as a low-resolution view of the anatomy. Because the ACS does not fully determine the anatomy, a conditional diffusion model (CARS) generates several plausible reference images consistent with it. A motion network (RAMI) compares each shot’s measurements against all references, weights the references with learned attention, and predicts per-shot motion; the training loss includes an ELBO term linking these weights to data consistency. The estimated motion initializes a motion-aware refinement and reconstruction. The method is evaluated with simulated motion on fastMRI FLAIR, low-field M4Raw, and a small in-house 0.55 T cohort, and compared against JSMoCo and an adapted MoPED.

Critical strengths

The problem they are addressing is a real and important application for MRI. Well motivated, well thought-out solutions for marginalizing the uncertainty of the motion and anatomy. Also their method does not require any extra information from a scout or navigator, which is important for medical applications.

Critical weaknesses
  1. All motions are simulated and assumed to be in-plane motions, which are unrealistic.
  2. The method shows only a modest ~10% improvements which is within the margin of their errors.
  3. Unpolished paper. Many mistakes across the whole paper. For example, the 9.65 db improvement claimed in the contribution statement does not appear anywhere in the paper.
  4. Runtime speed is not reported anywhere. Which is very important for medical applications.

That said, I'm not an expert in brain MRI or rigid-body motion modeling, so I may be miscalibrated.

Questions for authors

How does MARMI (K = 8) compare to using a single deterministic reference, such as the zero-filled ACS image? The current ablation only compares against one random sample with \lambda_ref = 0.

What's the runtime speed and how does it compare? Did you try on real patient motions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.