acceptodds
Withdrawn from ICLR 2027

Anatomy Is the Manifold: Learning Intrinsic Metrics Instead of Images for Prior-Guided MRI Reconstruction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Prior-guided multi-contrast MRI uses a fully sampled companion scan to reconstruct an undersampled target contrast. Networks that predict target intensities from the companion can transfer structures absent from the target and lose accuracy under distribution shift. We train a network to predict edge conductivities on an anatomy graph. These conductivities define a quadratic prior, and reconstruction solves a convex problem driven by target measurements. The formulation returns zero for zero measurements and provides an explicit expression for the effect of companion changes. We use held-out k-space to select among hand-crafted, learned and randomised metrics, supported by a risk-estimation result under uniform random sampling. In synthetic lesion experiments, learned metrics reduce the response to companion-only lesions while recovering more of target-only lesions than intensity priors. Learned geometry remains sensitive to companion appearance and registration. Randomising both during training reduces the PSNR loss from a one-pixel misalignment to 2.0 dB, compared with 3.6–5.0 dB for intensity priors trained with the same augmentation. Geometry selection remains within 0.6 dB of intensity selection in distribution and improves PSNR by 0.6–2.3 dB over the strongest intensity baseline on each cohort with a scanner, pathology, contrast or anatomy shift. On raw four-coil 0.3 T M4Raw data, the learned metric achieves higher PSNR than U-Net, MoDL and TGVN.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
MRI reconstruction, inverse problems, multi-contrast MRI, medical imaging, graph-based regularization, intrinsic geometry, distribution shift robustness, k-space model selection

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_HJweOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors propose a new method for prior-guided MRI reconstruction that learns edge conductivities from a companion scan. In prior-guided reconstruction, the authors ensure that zero-measurements yield a zero reconstruction, and perform randomization in the companion during training to improve robustness. The proposed method is shown on multiple different datasets and settings.

Critical strengths
  1. The paper addresses an important problem of prior-guided MRI reconstruction, which aims to improve accelerated MRI recovery by incorporating additional priors in the form of supplementary scans.
Critical weaknesses
  1. I think the authors use unclear and over-complicated terms and wording, making the paper difficult to follow or understand. While I have a knowledge on the topic, I had to re-read sentences and sections multiple times to understand aimed meaning. The clarity and readability of the paper must be significantly improved by using simpler language and more direct explanations.

  2. The experimental protocol has several issues. First, both IXI and BRATS datasets are single-coil datasets. This limits the representation of a realistic and practical MRI reconstruction scenario. Modern scanners use multiple coil channels, and the experiments should focus on multi-coil reconstructions. Second, the protocol for IXI and M4Raw are inconsistent. In IXI, the ratio between training and test sets is 540:9, whereas in M4Raw, it is 128:25. The test set in IXI should be expanded to enable a more robust evaluation. Third, baselines should reflect the current state-of-the-art. For instance, results are shown for compressed sensing TV (from 2007) and directional TV (from 2016), showing them on single-coil reconstruction, which is a limiting factor.

  3. Overall, the results section is not well structured. The goals and settings of the demonstrations are not sufficiently provided in advance. For instance, I had issues in understanding the purpose of the demonstrations in Section 4.1 and Section 4.2. It is also difficult to understand what conclusions should be drawn from them. No visual results are shown in the corresponding sections in the main paper. Table 1 and Table 2 report results, but there is no pattern in the conclusions. For each task, another column is shown to be the best.

  4. What's the volumetric results shown on IXI? Are they simply metrics calculated on the concatenation of the slices? Were there any volumetric trainings? If so, how is the reconstruction task built as a slice-wise reconstruction in the earlier experiments, and volumetric here? Similarly, how's the radial sampling simulated in IXI. The IXI dataset is a collection of single-channel MRI images, mostly containing T1, T2, and PD-weighted images. A simulation of a radial sampling on these images will not reflect a realistic reconstruction scenario. In order to claim demonstrations on radial sampling, it is required to use a radially acquired data.

  5. The comparison experiments should also include MRI contrast synthesis methods, as they are relevant benchmarks that only focus on generating target contrast images from given source contrast images.

  6. The reconstruction experiments are mainly shown for relatively low acceleration ratios of 4x and 8x. It would be helpful to show how the proposed method behave on even more aggressive acceleration, such as 12x or 16x.

Questions for authors
  1. The main experiments are focusing on single-coil reconstruction with a limited set of benchmark methods (not reflecting current state-of-the-art, such as diffusion models). Is the proposed approach demonstrated through large-scale experiments on multi-coil datasets containing different anatomies including brain, knee etc. and shown on more recent models? How does the proposed approach work on more aggressive acceleration ratios? What are the potential use cases of this approach? Are there any downstream evaluations?
Reviewer_v7aEOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a geometry-conditioned framework for prior-guided multi-contrast MRI reconstruction. Instead of directly predicting target image intensities from a fully sampled companion contrast, the method learns graph conductivities that define a spatially adaptive quadratic regularization prior. The target image is then reconstructed from undersampled k-space measurements through a convex optimization formulation.

The authors provide theoretical properties concerning the dependence of the reconstruction on target measurements and companion information. They also introduce contrast and alignment randomization to improve robustness, together with a held-out k-space selection strategy for choosing among learned and hand-crafted metrics.

Experiments on IXI, BraTS, prostate MRI, and raw multi-coil M4Raw data suggest improved robustness to distribution shifts and reduced sensitivity to structures appearing only in the companion image.

Critical strengths

1. A meaningful problem formulation with explicit constraints on information transfer.

The paper addresses a relevant concern in multi-contrast MRI reconstruction: anatomical information from a companion contrast may introduce misleading structures into the reconstructed target. Restricting the companion to defining spatial regularization, rather than directly predicting target intensities, is an interesting approach. The theoretical analysis and controlled lesion experiments provide useful insights into the behavior of this formulation.

2. Extensive robustness evaluation.

The authors evaluate distribution shifts involving scanner, anatomy, target contrast, and companion misalignment. The inclusion of controlled lesion experiments, comparisons with multiple reconstruction methods, and an additional raw multi-coil MRI experiment strengthens the empirical evaluation. The held-out k-space selection strategy is also practically interesting.

Critical weaknesses

1. The relationship between the theoretical formulation and the actual reconstruction procedure is not fully convincing.

The paper emphasizes the properties of the exact solution to a convex quadratic optimization problem. However, the implemented reconstruction uses a fixed number of conjugate-gradient iterations.

More importantly, Appendix O shows that further convergence actually reduces reconstruction quality. Increasing CG iterations from 60 to 300 decreases PSNR from 35.45 dB to 34.35 dB, despite substantially reducing the optimization residual.

This raises an important question about how much of the reported performance comes from the proposed metric and how much depends on early stopping. Several theoretical interpretations concern the converged solution, whereas the reported experimental results rely on a truncated solver. Although this does not invalidate the theoretical propositions, it weakens their direct connection to the practical method.

2. The novelty and practical advantages over existing approaches need stronger justification.

The proposed formulation builds upon established graph Laplacian regularization, learned spatial regularization weights, and measurement-based model selection. While their combination is useful, the main methodological novelty is not entirely clear.

Furthermore, the improvements under distribution shift appear to rely substantially on classical hand-crafted metrics and test-time selection. For example, Table 9 shows that hand-crafted metrics are selected for the majority of BraTS samples. This makes it difficult to determine whether the learned metric itself is responsible for the reported generalization improvements.

I also find the comparison with modern generative reconstruction methods incomplete. The paper mentions diffusion-based MRI reconstruction with side information in the related work, but does not include a diffusion-based baseline. Such methods are relevant alternatives for modeling complex image distributions and addressing image smoothing, although they may also introduce hallucinations.

Finally, most cross-domain experiments rely on simulated k-space from magnitude images. While the raw MRI experiment is valuable, its improvements over the strongest baselines are relatively modest. The practical advantages under more realistic acquisition conditions therefore remain somewhat uncertain.

Questions for authors
  1. Theoretical formulation versus early stopping: How should we interpret the theoretical guarantees when the best reconstruction performance is obtained before CG convergence? Do the conclusions concerning companion-induced changes and lesion responses remain quantitatively consistent when the solver approaches convergence?

  2. Contribution of learned metrics: Since the held-out selection strategy frequently chooses hand-crafted metrics on BraTS, could the authors clarify how much of the cross-domain improvement can be attributed specifically to learning the metric, rather than classical regularization and model selection?

  3. Comparison with diffusion-based reconstruction: The paper discusses diffusion-based MRI reconstruction with side information in the related work, but does not evaluate such methods. Could the authors explain this omission and clarify how the proposed method compares conceptually or empirically with data-consistent diffusion-based reconstruction? This would help establish its advantages beyond conventional intensity-prediction baselines.

Reviewer_eZ9XOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This manuscript propose to learn a metric graph for for multi-contrast MRI translation and reconstruction. The pipeline requires a fully sampled companion from the target in order to learn and infer the graph refinements. Those converged graphs define the metric for the MRI reconstruction optimization steps. This should improve the reconstruction quality as shown on IXI, BRATS and ADC MRI datasets.

Critical strengths

Learning the metric over the anatomy underlying the image makes sense as an idea to improve reconstruction quality in MRI translation.

Critical weaknesses
  • The problem scope is small and conditions are demanding ( requiring accurate companion of the MRI and an anatomy prior ). This makes ICLR not the proper venue for this manuscript, but more of MICCAI , or other medical imaging journals.

  • The generalization of the followed approach of learning a U-Net over anatomy graph model is questionable. Given that the training happen in more complex pipeline that involved mid-step optimization with conjugate gradients.

  • The empirical results in Table 1 and 2 show minimal to no improvement over the baselines despite the very complex pipeline followed. No reporting of randomness in the results makes the judgment of the presented results very difficult.

  • The presentation can be improved. The writing of the paper is LLM-looking, most sentences are very long with colon (:) , seemi colon (;) and em dashes (---) in the middle making it hard to understand. The figures and plots in the paper are not informative either (se Figure 2).

Questions for authors

how do you justify the complexity of the proposed pipeline ( with its cost not reported ) compared to the other baselines in Table 1 and 2 compared to the gains in the PSNR/SSIM ?

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

Related papers

Discussion (0)

Sign in to comment.