acceptodds
Withdrawn from ICLR 2027

Covariant Wavelet Scattering for Robust MRI Reconstruction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Unrolled variational networks are the dominant framework for accelerated MRI reconstruction, but their learned regularizers can overfit the sampling pattern seen in training. We replace the learned spatial operator of the regularizer with a fixed covariant wavelet scattering transform: a low-pass filter and oriented Morlet wavelets followed by a modulus, without the averaging that makes standard scattering invariant. The resulting operator tiles the full frequency plane with explicit frame bounds, and we prove it is translation-equivariant, retains every frequency band through a lower frame bound, and is non-expansive and stable to small deformations. The filterbank itself has no trainable parameters. On fastMRI knee and brain data, AnalyticVarNet improves over a neural-operator baseline in-distribution and is substantially more robust to unseen incoherent sampling masks, with fewer parameters.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
MRI denoising.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3RGZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AnalyticVarNet, an unrolled framework for multi-coil MRI reconstruction that incorporates a fixed covariant wavelet scattering operator. The operator combines Gaussian low-pass filtering with the modulus of oriented Morlet wavelet responses. It omits subsequent averaging and subsampling to preserve spatial information, while learned layers predict reconstruction corrections. The authors analyze the frequency coverage, translation equivariance, and stability of the fixed operator and evaluate the model on fastMRI knee and brain data. The main empirical advantages are improved robustness to unseen sampling patterns and reduced parameter count and inference time relative to the NO baseline.

Critical strengths
  • Relevant problem and meaningful generalization setting. Robustness to acquisition-pattern changes is an important issue in learned MRI reconstruction.
  • A well-motivated integration of analytic and learned components Combining fixed frequency and orientation analysis with learned adaptation and explicit data consistency is a well-motivated approach to reconstruction. The contribution lies mainly in this task-specific integration.
  • Useful efficiency gains. The model reduces trainable parameters from approximately 141.8M to 57.4M and lowers reported inference time from 177.38 to 145.34 ms per slice.
Critical weaknesses
  • Core attribution is unproven: only one baseline (NO), differing simultaneously in width, capacity, and structure. A matched-width, matched-parameter standard-conv ablation is missing.
  • Robustness rests on a single baseline (NO); E2E-VN and diffusion methods are never tested for robustness. NO's collapse to ~11–13 dB/SSIM 0.34 on Gaussian/Poisson despite data consistency is suspicious and should be checked for evaluation artifacts.
  • Only J=1 Morlet bank is used; genuine scattering multi-scale structure is absent. The "scattering" claim in the title/abstract is overstated.
Questions for authors
  • Given your claim that finer low-frequency partitioning matters at higher native resolution, is J=1 validated on 640×640 data? Can you provide an L (number of orientations) sensitivity analysis?
  • Does the positive lower frame bound apply to discrete Nyquist grids or a specified band-limited space? A finite collection of decaying filters cannot have a uniform positive lower bound over all continuous frequencies. Are the reported AA and BB established analytically or evaluated numerically?
  • Are standard deviations computed across slices or volumes? Are training, fine-tuning, and evaluation sets disjoint at the subject level? Can volume-level paired differences or confidence intervals be reported?
Reviewer_zqTtOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AnalyticVarNet, an unrolled variational network for accelerated multi-coil MRI reconstruction that incorporates a fixed covariant wavelet scattering transform. The authors argue that learned spectral operators may overfit to the sampling patterns encountered during training, potentially limiting generalization to unseen acquisition settings.

To address this issue, the proposed method replaces learned spectral filtering with a fixed Morlet wavelet filterbank consisting of a low-pass component and oriented wavelets followed by a modulus operation. Unlike conventional scattering representations, the method removes spatial averaging and subsampling to preserve translation equivariance and spatial localization. The authors establish frame bounds, translation equivariance, non-expansiveness, and deformation stability for the resulting scattering operator.

The proposed representation is incorporated into an unrolled MRI reconstruction framework with learned regularization and data consistency. Experiments on fastMRI knee and brain datasets show modest improvements over a neural-operator baseline under in-distribution evaluation and considerably larger gains under certain unseen sampling masks and acceleration factors. The method also reduces the number of trainable parameters and inference time.

Critical strengths

1. A structured and interpretable spectral representation for MRI reconstruction.

The paper introduces a fixed Morlet wavelet representation into an unrolled reconstruction framework, providing an explicit spectral decomposition rather than learning the entire filterbank from data.

The decision to omit scattering averaging and subsampling is reasonable for image reconstruction, where spatial localization must be preserved. The associated mathematical analysis provides useful guarantees concerning the fixed operator's frequency coverage, equivariance, and stability.

Although these properties largely follow from classical wavelet and scattering theory, their integration into a reconstruction architecture provides a meaningful inductive bias.

2. Promising robustness under sampling distribution shifts with reduced model complexity.

The experimental results demonstrate substantial improvements over the neural-operator baseline under several unseen sampling patterns.

For example, Table 2 reports zero-shot PSNR improvements of approximately 9.8 dB, 14.8 dB, and 14.5 dB on the fastMRI knee dataset for radial, Gaussian, and Poisson sampling patterns, respectively. The authors also provide additional cross-mask and cross-acceleration experiments on brain data.

Furthermore, the proposed model uses approximately 2.47 times fewer trainable parameters and achieves an approximately 18% reduction in inference time compared with the neural-operator baseline.

These findings suggest that analytically structured filtering can be practically beneficial, particularly when training data cover only a limited range of acquisition settings.

Critical weaknesses

1. Insufficient methodological justification and limited novelty of the proposed covariant scattering representation.

The central motivation is that learned spectral operators may overfit to the training sampling pattern, whereas a fixed Morlet wavelet filterbank provides complete spectral coverage and potentially improves generalization.

While this motivation is plausible, the paper does not sufficiently establish why the specific covariant Morlet scattering representation is necessary or preferable to simpler alternatives.

In particular, complete frequency coverage is not unique to Morlet wavelets. Other fixed transforms, such as Fourier, DCT, or conventional wavelet decompositions, can also provide structured spectral coverage without learning the underlying filters. Similarly, fixed Gabor-like filters or appropriately constrained learned filterbanks could potentially provide comparable spectral properties.

The paper does not compare the proposed method against these alternatives. Consequently, it is unclear whether the observed benefits arise specifically from the Morlet scattering representation, from using a fixed filterbank, or more generally from reducing the flexibility of the learned spectral operator.

The methodological novelty is also somewhat limited. The proposed representation consists of standard wavelet convolutions followed by modulus operations, with the averaging and subsampling of conventional scattering removed to retain spatial localization. Translation equivariance and the relevant stability properties are largely inherited from classical convolutional wavelet representations.

Furthermore, the paper does not directly demonstrate the hypothesized failure mechanism of learned spectral operators. Although the authors argue that learned kernels may suppress frequency bands that are underrepresented during training, no analysis of the learned baseline's spectral response is provided to establish that this actually occurs.

This weakness is consequential because the main methodological contribution relies on the premise that the proposed analytic representation addresses a specific limitation of learned neural operators. The experiments show that the resulting architecture can perform well, but they do not establish that the particular choice of covariant wavelet scattering is necessary or uniquely advantageous.

2. Insufficient isolation of the proposed mechanism and limited connection between theoretical guarantees and empirical robustness.

The central theoretical argument is that the fixed scattering operator provides frequency coverage and stability guarantees independent of the training distribution.

However, Proposition 1 concerns only the fixed scattering operator, not the complete reconstruction network. The subsequent learned convolutional layers, nonlinearities, normalization, and reconstruction cascades are not covered by these guarantees.

In particular, a lower frame bound for the fixed transform ensures that the transform does not eliminate the input's overall energy, but it does not establish that all frequency information remains recoverable after the modulus operation and subsequent learned processing. Nor does it imply robustness of the complete reconstruction network to changes in the sampling operator.

Therefore, the proposed explanation for the improved out-of-distribution performance remains largely indirect.

The experimental ablations also do not sufficiently isolate the source of improvement.

First, AnalyticVarNet has substantially fewer trainable parameters than the neural-operator baseline. Without a parameter-matched baseline, it is difficult to distinguish the benefits of the proposed spectral representation from those of reduced model capacity and potentially reduced overfitting.

Second, the method appears to introduce scattering-based processing beyond the main cascade regularizer. The initialization module follows the regularizer architecture, while Supplementary Appendix B explicitly states that the sensitivity estimation network also uses scattering blocks. Consequently, the reported improvements may arise from changes to multiple reconstruction components rather than specifically from the proposed regularizer.

Third, the reported robustness improvements vary considerably across datasets. The extremely large improvements under radial, Gaussian, and Poisson masks on the knee dataset contrast with substantially smaller improvements under corresponding masks on the brain dataset. While this does not invalidate the results, the paper does not sufficiently investigate why the same proposed mechanism produces such different gains.

Finally, the zero-shot cross-resolution results require a more nuanced interpretation. The large PSNR improvement under interpolation-based transfer does not persist under native-resolution evaluation, where the improvement over the neural-operator baseline is only 0.23 dB in PSNR.

These concerns are important because the empirical results are presented as evidence that analytically guaranteed spectral coverage improves robustness. Without stronger component-wise and capacity-controlled comparisons, the experiments do not convincingly establish this causal connection.

Questions for authors

1. Motivation for covariant Morlet scattering (clarification and additional evidence).

The paper motivates the proposed representation by arguing that fixed spectral coverage helps prevent overfitting to training sampling patterns.

However, several simpler transforms can also provide complete or well-controlled frequency coverage.

What specific properties of the Morlet scattering representation make it more suitable for MRI reconstruction than conventional fixed wavelet transforms, Fourier-based features, or other fixed oriented filterbanks?

Could the authors provide controlled comparisons with at least one alternative fixed spectral representation under the same reconstruction architecture?

Such comparisons would help establish whether the proposed scattering representation offers benefits beyond the general use of fixed spectral filtering.

2. Evidence for the claimed spectral coverage mechanism (clarification).

The authors suggest that learned neural operators may suppress frequency bands that are uninformative under the training sampling pattern, whereas the fixed scattering filterbank prevents this behavior.

Is there direct empirical evidence that the learned neural-operator baseline actually exhibits such frequency suppression?

For example, analyzing the learned filter responses or reconstruction errors across frequency bands could help determine whether the observed robustness improvements are consistent with the proposed explanation.

Additionally, since Proposition 1 applies only to the fixed scattering operator, how should its frame bounds and deformation stability be interpreted in relation to the robustness of the complete learned reconstruction network?

3. Isolating the contribution of the scattering regularizer (additional evidence).

The proposed model has substantially fewer trainable parameters than the neural-operator baseline, and scattering-based blocks are also used in the sensitivity estimation and initialization components.

Could the authors provide component-wise ablations that replace scattering independently in the main regularizer, sensitivity estimation network, and initialization module?

A parameter-matched neural-operator or CNN baseline would also help distinguish the effects of the proposed inductive bias from those of reduced model capacity.

These experiments are important for establishing which architectural changes are responsible for the reported improvements.

4. Generalization claims and evaluation reliability (clarification and additional evidence).

The reported zero-shot cross-mask gains are particularly large on the knee dataset but substantially smaller on the brain dataset.

Can the authors explain this difference and clarify whether the neural-operator baseline was evaluated under exactly matched mask handling, data preprocessing, and normalization procedures?

Furthermore, the in-distribution PSNR improvements are relatively small compared with the reported standard deviations. Could the authors provide results across multiple training seeds or paired statistical comparisons?

Finally, could the authors clarify why the large improvement in interpolation-based cross-resolution transfer does not translate into a similarly substantial PSNR gain under native-resolution evaluation?

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

This paper proposes AnalyticVarNet, which incorporates a fixed covariant wavelet scattering operator into the regularization module of a variational network to improve the generalization of MRI reconstruction to unseen sampling patterns. The authors analyze the properties of the proposed operator in terms of frequency coverage, translation equivariance, and deformation stability, and evaluate the method on the fastMRI knee and brain datasets.

Critical strengths

The paper addresses a practically relevant problem, and the proposed method has a clear structure. The combination of fixed frequency representations and learnable feature extraction is reasonably motivated. The reported experimental results show substantial improvements in reconstruction performance under certain unseen sampling patterns, while also reducing the number of trainable parameters and improving inference speed.

Critical weaknesses
  1. The reproducibility of the core cross-sampling experiments requires further verification.

The paper reports substantial performance improvements under Radial, Gaussian, and Poisson sampling patterns. However, the NO baseline achieves PSNR values of only 11.85 dB and 13.16 dB under Gaussian and Poisson sampling, respectively. Given the magnitude of the reported improvements, the paper's central claims depend heavily on the baseline performance. Although the authors state that they used the official implementation and identical training configurations, insufficient evidence is currently provided to establish consistency in training settings, actual sampling ratios, calibration regions, and evaluation protocols. The authors should provide complete configurations, mask generation procedures, and reproducible evaluation results. The current evidence is insufficient to conclude that the baseline was incorrectly reproduced, but this issue requires a thorough explanation.

  1. The connection between the theoretical properties and the observed generalization gains is insufficiently established.

The authors establish theoretical properties of the fixed scattering operator, including a lower bound on frequency coverage. However, these guarantees apply to the operator itself and do not directly imply that the complete network will necessarily exhibit stronger out-of-distribution generalization. Furthermore, the initialization module, coil sensitivity estimation module, and main regularization network all involve architectural modifications, making it difficult to isolate the contributions of individual components. The authors should consider adding parameter-matched learnable convolutional baselines, targeted ablation studies, and cross-sampling evaluations on additional datasets to provide more convincing evidence for the paper's central claims.

Questions for authors

see weaknesses

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

Related papers

Discussion (0)

Sign in to comment.