acceptodds
Under review as a conference paper at ICLR 2027

UEPS: Robust and Efficient MRI Reconstruction

Abstract

Deep unrolled models (DUM) have become the state-of-the-art for accelerated MRI reconstruction, yet their robustness under domain shift remains a critical barrier to clinical adoption. In this work, we identify coil sensitivity map (CSM) estimation as the primary bottleneck limiting generalization. To address this, we propose UEPS, a novel DUM architecture featuring three key designs: (i) an Un-rolled Expanded (UE) design that eliminates CSM dependency by reconstructing each coil independently; (ii) progressive resolution, which leverages k-space-to- image mapping for efficient coarse-to-fine refinement; and (iii) sparse attention tailored to MRI’s 1D undersampling nature. These physics-grounded designs en- able gains in robustness and computational efficiency. We construct a large-scale zero-shot transfer benchmark comprising 10 out-of-distribution (OOD) test sets spanning diverse clinical shifts—anatomy, view, contrast, vendor, field strength, and coil configurations. Extensive experiments demonstrate that UEPS consis- tently and substantially outperforms existing DUM, end-to-end, diffusion, and un- trained methods across all OOD tests, achieving state-of-the-art robustness with low-latency inference suitable for real-time deployment.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Robustness, Accelerated MRI, Deep unrolled models, Sparse attention

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_qzS2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper identifies coil sensitivity map (CSM) estimation in unrolled networks as a bottleneck for generalization and proposes a new architecture, UEPS, that bypasses learned CSM estimation. The authors evaluate the approach on several out-of-distribution (OOD) datasets and shows that UEPS improves performance over unrolled networks that rely on learned sensitivity maps.

Critical strengths
  • Identifying that learned sensitivity maps contribute to OOD robustness degradation.
  • Computationally efficient application of vision transformers to coil images.
  • Extensive OOD evaluation across different datasets.
Critical weaknesses
  • Currently, many design choices are changed, hence it is not clear whether the problem is really the estimated sensitivity maps. The experiments do not sufficiently control for that.
  • It is unclear how much of the performance improvement comes from the transformer-based architecture rather than the proposed coil-wise reconstruction approach. A more controlled ablation would be useful.
  • Section 4.2 shows that learned sensitivity maps are affected by distribution shifts. However, the question that the authors initially raise is how learned sensitivity maps in combination with learned reconstruction affect the final reconstruction.
  • The experiments do not consider training on diverse datasets, which has been shown to substantially improve robustness. It is unclear whether UEPS offers additional benefits in this setting.
  • The authors consider only FastMRI brain as training distribution.
Questions for authors

Main questions

  • Can the authors provide a model trained on all datasets to see how large the gap is with respect to an ideally robust model? How does UEPS compare to simply mixing datasets? Are there additional benefits of UEPS when the model is trained on diverse datasets?

  • “Our analysis reveals that learned CSM estimation becomes less reliable under domain shift.” What about training and applying models with precomputed ESPIRiT CSMs? In this case, CSM estimation does not depend on the learned CSM model in an end-to-end network. Are models also less robust in this case? I realize that in this case one would not use RSS images as training and evaluation references but rather MVUE targets based on the sensitivity maps. Hence, my question is specific to this setup.

  • Learning in a coil-combined image space has the benefit that measurements across coils can be fused to provide more information. How does UEPS perform at higher acceleration factors?

  • “An additional benefit is the implicit curriculum learning.” How is the curriculum chosen? How much does it affect training in terms of accuracy and training time?

  • Can the authors perform an ablation using simple U-Nets? How much performance does this transformer-based approach really add compared to a simpler U-Net-based approach?

  • It would make sense to have an E2E VarNet ablation where the sensitivity model is omitted and reconstruction is done using the authors' proposed approach of reconstructing individual coil images, but with E2E VarNet U-Nets. This would directly test the authors' hypothesis of sensitivity maps as robustness degraders and should serve as an additional baseline to check (i) whether sensitivity maps really are a problem and (ii) whether the transformer-based design choices provide additional benefits.

  • How does the model perform under coil shifts?

  • How are the end-to-end VarNets trained? Lin and Heckel (2024) show that when VarNet is trained on cropped images, an artificial distribution shift due to different image sizes can occur, which needs to be corrected, or VarNet needs to be trained on uncropped images. They also show that VarNet's PSNR and SSIM scores can be heavily affected by slightly incorrect, visually unrecognizable global mean and variance predictions under distribution shifts. Do the authors consider these factors in their evaluation?

  • Why was FastMRI Brain chosen for training?

Minor Comments

  • Some terms like “authentic high-frequency information” or “our design inherently strengthens the progressive refinement nature of DUM” are abstract and difficult to understand. I suggest omitting or replacing them with concrete descriptions.

  • Related work can focus more on approaches that tackle the same problem as the paper, which is robustness for accelerated MRI.

  • “Our progressive resolution design is, to the best of our knowledge, the first design to unify cascaded refinement and resolution scheduling into a single, up-sampling-only loop, which substantially improves efficiency by operating at lower resolutions for the bulk of the computation.” The way this is formulated makes it sound like a core benefit. I agree that this part is smart, but the engineering stems from a compromise that ideally would not be needed, rather than an addition that would definitely improve performance.

  • Figure 3 is uninformative since it is not clear how the attention map is computed.

  • Consider capitalizing Figure, Table, and Section references.

  • Reconstruction figures are way too small. The space could be better used. Error maps are not really needed. This would save space, and reconstruction artifacts could be directly pointed out in the reconstruction images.

  • “With the RSS combination, which an ideal CSM should recover.” RSS is not recovered by ideal CSMs due noise and the assumptions made by the considered forward model.

  • References in Figure 1 need to be fixed.

  • References need to be checked again. For example, SwinIR was published as workshop paper in IEEE/CVF International Conference on Computer Vision Workshops and the Flex Attention reference seems to have an additional incorrect "2(3)" field. Also, some references refer to arXiv versions when peer-reviewed articles are already available.

References

Lin, Kang, and Reinhard Heckel. "Robustness of Deep Learning for Accelerated MRI: Benefits of Diverse Training Data". ICML, 2024.

3 comments
AuthorsOpenReview ↗

Response to Reviewer qzS2 Part (1/2)

We thank you for your detailed and constructive feedback. Below, we summarize each concern and respond point by point, grouping related weaknesses (W), questions (Q), and minor comments (M).

W1–W2 & Q5–Q6: Controlled ablations of coil-wise reconstruction and backbone choice.

A1: Thanks for suggesting these controlled comparisons. We have expanded Table 3 with a more controlled ablation study, comparing E2E-VarNet, U-Net-based UE, ViT-based UE, and variants with progressive resolution(PR) and sparse attention(SA). All configurations, including the E2E-VarNet ablation baseline, use the same training protocol for fair comparison. The average OOD results are:

ModelOOD PSNR (dB) ↑OOD SSIM (%) ↑Runtime (ms) ↓
E2E-VarNet36.2491.4321.78
UE (U-Net)37.1092.48155.76
UE (ViT)37.5392.7997.17
UE (ViT) + PR37.5992.9642.31
UE (ViT) + SA37.8393.2181.24
UEPS37.7593.1138.01

As expected, UE yields the largest improvement in average OOD reconstruction quality due to its coil-wise reconstruction design, which avoids reliance on potentially unreliable CSM estimates under distribution shifts, while the ViT backbone, progressive resolution, and sparse attention provide additional gains. Although UE increases runtime, progressive resolution and sparse attention substantially reduce the computational cost of coil-wise reconstruction.


W3 & Q2: How learned CSMs affect final reconstruction; training and testing with precomputed ESPIRiT maps and CSM-defined MVUE targets.

A2: We trained and evaluated an E2E-VarNet variant using precomputed BART/ESPIRiT maps throughout both training and testing, with separate models for the RSS and MVUE reference settings. For MVUE, we used CSM-based output combination and the corresponding fully sampled reference consistently during training and evaluation. All models use identical reconstruction configurations and training protocols, differing only in their CSM-related settings. The average results across the ten OOD conditions are:

ModelOOD PSNR (dB) ↑OOD SSIM (%) ↑
E2E-VarNet (RSS)36.2491.43
E2E-VarNet with fixed BART CSMs (RSS)35.8890.77
E2E-VarNet with fixed BART CSMs (MVUE)36.4291.56
UE-U-Net (RSS)37.1092.48

Under the RSS protocol, the fixed-BART-CSM model remains 1.22 dB below UE-U-Net, showing that precomputed CSMs alone do not eliminate the OOD reconstruction gap. The fixed-CSM pipeline is not fully end-to-end: CSM estimation is performed externally and is not jointly optimized with the image-refinement modules. Under domain shift, these modules can encounter unexpected variations in the CSM-dependent image representations. UE eliminates this dependence through independent coil-wise reconstruction and achieves the best OOD reconstruction performance.


W4–W5 & Q1, Q9: Mixed-dataset training, and the choice of fastMRI Brain as the training source.

A3: Thank you for highlighting diverse-data training. We recognize its benefits for OOD robustness, as demonstrated by prior work (Lin & Heckel, 2024). For our main benchmark, we made a trade-off between training-data diversity and OOD evaluation coverage. Public fully sampled multi-coil MRI datasets are limited, and adding more sources to training leaves fewer acquisition domains entirely unseen during evaluation. We therefore chose fastMRI Brain, the largest dataset in our collection, as the sole training source and reserved the other datasets for OOD testing, enabling architectural comparisons under a common training distribution.


Q3: Performance at higher acceleration factors

A4: Prior work has demonstrated benefits from training across multiple acceleration factors (Lin & Heckel, 2024). Such sampling variations can be retrospectively generated from fully sampled k-space and incorporated during training. To focus on cross-dataset robustness across vendors, field strengths, anatomies, and coil configurations. we use a common protocol with nominal 4× acceleration and fixed input resolution. Extending our evaluation to higher and unseen acceleration factors remains an important direction for future research.

AuthorsOpenReview ↗

Response to Reviewer qzS2 Part (2/2)

Q4 & M1, M3: The motivation of progressive resolution.

A5: Progressive resolution is designed primarily to improve inference efficiency. Our experimental results (Table 3) show that this design substantially reduces runtime while maintaining comparable OOD reconstruction quality. We have streamlined the manuscript description to focus on this efficiency benefit with concrete evidence.


Q7: Generalization under coil shifts.

A6: UEPS treats coils as a batch dimension with shared weights, supporting variable coil counts. The fastMRI Brain training source contains 2–28 coils from Siemens acquisitions at 1.5T/3T. Our benchmark evaluates transfer across brain and mixed-anatomy acquisitions. Representative coil configurations and RSS-reference PSNR results are shown below:

DatasetCoilsE2E-VarNet PSNR (dB)UEPS PSNR (dB)Gain (dB)
M4Raw GRE435.6237.47+1.85
AHEAD, mean over five echoes3235.3338.46+3.13
Stanford2D3–3238.5440.52+1.98

Across all ten OOD conditions, UEPS improves PSNR over E2E-VarNet by 1.45–3.53 dB, including acquisitions with diverse coil configurations.


Q8: How the VarNets are trained

A7: For the benchmark results shown in Figure 1 and Table 8, we used the official fastMRI E2E-VarNet implementation, with model inputs standardized to 320×320 during both training and testing.

For the ablation studies, we used our own reimplementation of E2E-VarNet and trained it using the same training pipeline as the other ablation configurations, ensuring a fair comparison of individual components.


M2: A related-work discussion more focused on robustness in accelerated MRI.

A8: We have expanded the Related Work section to better summarize progress in robust accelerated MRI reconstruction.


M4: How the attention maps in Figure 3 are computed.

A9: We have added a description of the attention-map computation in Section 3.4. Specifically, the maps are extracted from the full-attention ViT-based UE model. For each selected query patch, we extract its softmax attention weights over all key patches and average them across all cascades, Transformer layers, and attention heads, separately for each coil. We then reshape the resulting vector into the two-dimensional patch grid, apply min–max normalization, and resize it to the image resolution for visualization.


M5–M9: Wording, formatting, figure layouts, citations, and bibliographic entries.

A10: Thanks for these suggestions and corrections. We have revised the relevant wording, formatting, figure layouts, citations, and bibliographic entries in the manuscript accordingly.


References

Lin, Kang, and Reinhard Heckel. "Robustness of Deep Learning for Accelerated MRI: Benefits of Diverse Training Data". ICML, 2024.

Thank you for the additional experiments and clarifications. The U-Net ablation addresses my concern about separating the benefit of coil-wise reconstruction from the transformer backbone. The fixed-ESPIRiT experiments also strengthen the study. However, a major concern about the scope and significance of the robustness claim remains. I therefore maintain my score of 2 (weak rejection).

The results demonstrate improvements across several OOD test conditions after training on fastMRI Brain, but they do not establish how these improvements depend on the diversity of the training distribution. UEPS may retain its advantage under broader training, or it may primarily compensate for limitations of single-source training. Both would be useful findings, but they imply different conclusions about the scope and practical significance of the proposed robustness improvement.

In my view, a more systematic analysis of how much and when CSM dependence limits generalization, particularly whether this limitation persists with more diverse training data, would strengthen the central scientific contribution. The transformer and efficiency components could then provide a practical implementation of the resulting reconstruction approach.

Reviewer_C7aZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the robustness and computational cost of multi-coil MRI reconstruction under distribution shift. It avoids explicit coil sensitivity maps and combines progressive spatial resolution with sparse attention. The experiments cover in-distribution reconstruction, multiple external test sets, and inference efficiency. I view its main contribution as a targeted system design supported by cross-domain experiments.

Critical strengths
  1. The cross-domain gains are supported by substantive evidence. In Table 2 on page 7, UE reaches an average OOD PSNR of 37.54 dB, compared with 35.33 dB for E2E-VarNet, while their in-distribution results are similar. In Table 8 on page 16, the full UEPS model has the highest reported PSNR on all ten OOD test sets. Several sets share a data source, which limits their independence, but the breadth of evaluation remains a strength.
  2. The efficiency study addresses a practical design objective. Table 3 and Appendix B.3 compare design choices with similar parameter counts and cascade settings. The reported inference time falls from 97.28 to 38.02, corresponding to roughly a 2.56-fold speedup. Evaluating reconstruction quality together with inference cost adds value. These results concern inference, not training speed.
Critical weaknesses
  1. The design is effective, but the central mechanistic explanation is stronger than the evidence. The experiments do not yet establish sensitivity-map errors as the main cause of cross-domain degradation. Image-intensity or RSS-based proxies are not direct measurements of sensitivity-map error. Changing the reconstruction representation also changes the mapping learned by the network. Earlier work such as GrappaNet and NoSENSE already explored reconstruction without explicit sensitivity maps. The paper should therefore distinguish its combined design and empirical validation from a new explanation of domain-shift failures. The evidence currently supports the former more strongly.
  2. The aggregation rules and comparison protocols affect the precise scope of the overall advantage. The ten OOD test sets come from five data sources, and several AHEAD echoes share subjects. They should not be interpreted as ten independent external validations. The UEPS point in Figure 1 on page 2 is approximately 38.13, whereas the mean of the ten OOD PSNR values in Table 8 is about 37.75. Including the in-distribution result gives approximately 38.13. This suggests that the plot may include the in-distribution result, which needs clarification. DDS also uses a different sampling scheme and reference image, so the overall quality–efficiency ranking should distinguish protocols. These issues affect the magnitude and scope of the reported advantage. The recalculations do not reverse the main improvement.
Questions for authors

1.Please specify the test sets and aggregation rule behind Figure 1, including whether it contains the in-distribution result. Do the conclusions change when results are aggregated by independent data source? 2.Is there controlled evidence separating sensitivity-map errors from changes in representation and network design? If not, can the mechanistic explanation be narrowed? 3.Which update order is used in the implementation corresponding to Equations (5)–(9)? Does correcting the equivalence affect any experiment?

2 comments
AuthorsOpenReview ↗

Response to Reviewer C7aZ

We thank you for your insightful feedback. Below, we summarize each concern and respond point by point, grouping related weaknesses (W) and questions (Q).

W1 & Q2: Component contributions.

A1: To better distinguish the contributions of individual components from the overall effect of their combined design, we expanded the ablation study in Table 3 to include E2E-VarNet, U-Net-based UE, ViT-based UE, and variants incorporating progressive resolution (PR) and sparse attention (SA). All configurations, including the E2E-VarNet ablation baseline, follow the same training protocol. Note that the E2E-VarNet results are from our reimplementation of the original method (Sriram et al., 2020a), providing a stronger baseline than the original version used in the benchmark. The results below are averaged over the OOD test sets only.

ModelOOD PSNR (dB) ↑OOD SSIM (%) ↑Runtime (ms) ↓
E2E-VarNet36.2491.4321.78
UE (U-Net)37.1092.48155.76
UE (ViT)37.5392.7997.17
UE (ViT) + PR37.5992.9642.31
UE (ViT) + SA37.8393.2181.24
UEPS37.7593.1138.01

The largest observed stepwise improvement in OOD reconstruction quality comes from moving from E2E-VarNet to U-Net-based UE, indicating that coil-wise reconstruction without reliance on explicitly estimated CSMs can improve robustness under domain shifts. The ViT backbone, PR, and SA primarily improve inference efficiency while maintaining competitive reconstruction quality. Together, they reduce average runtime from 155.76 to 38.01 ms, while further improving OOD PSNR to 37.75 dB.


W2(a) & Q1: The test sets and aggregation rule behind Figure 1

A2: For all methods in Figure 1, the originally reported averages included both the ID and OOD results. We have updated the manuscript and now report OOD-only averages consistently across all figures and tables, explicitly stating that Average refers to the mean across OOD datasets.

Under this OOD-only aggregation rule, the paper’s main conclusion remains unchanged: UEPS substantially outperforms the compared methods in average OOD reconstruction quality. For example, the table below reports the average OOD performance of UEPS and E2E-VarNet.

ModelOOD PSNR (dB) ↑OOD SSIM (%) ↑
E2E-VarNet35.3390.70
UEPS37.7593.11

UEPS outperforms E2E-VarNet by 2.42 dB in PSNR and 2.41 percentage points in SSIM.


W2(b): Independence of the OOD test sets and DDS protocol.

A3: The benchmark contains ten OOD test conditions from five data-source groups. For example, the five AHEAD echoes (Alkemade et al., 2020) provide images with different tissue contrasts and signal intensities arising from different inversion/echo conditions. Evaluating them separately allows us to examine reconstruction performance across contrast-related changes in image appearance. The two CC-359 views (Souza et al., 2018), axial and sagittal, present different cross-sections of the same 3D brain volumes, with different in-plane arrangements of anatomical structures. They therefore allow us to assess view-related changes in anatomical appearance and spatial structure. These conditions share subjects within each source. We report these conditions separately to provide a more detailed assessment of contrast- and view-related distribution shifts.

To better assess DDS’s actual reconstruction capabilities, we use the officially released pretrained weights and follow its original evaluation protocol, which uses Gaussian sampling and an MVUE reference (Chung et al., 2024).


Q3: The implemented update order corresponding to Equations (5)–(9)

A4: Equations (5)–(9) describe the cascade update implemented in our Unroll model. Its overall structure follows the common DUM framework of alternating learned image refinement and measurement-based data consistency. Within each cascade, the implemented order is network refinement → Fourier transform → data consistency → inverse Fourier transform. Our CSM-related ablations follow the same Unroll update scheme, providing a consistent basis for comparison.


References

Sriram, Anuroop, et al. "End-to-End Variational Networks for Accelerated MRI Reconstruction". MICCAI, 2020a.

Alkemade, Anneke, et al. "The Amsterdam Ultra-high field adult lifespan database (AHEAD): A freely available multimodal 7 Tesla submillimeter magnetic resonance imaging database". NeuroImage, 2020.

Souza, Roberto, et al. "An open, multi-vendor, multi-field-strength brain MR dataset and analysis of publicly available skull stripping methods agreement". NeuroImage, 2018.

Chung, Hyungjin, Suhyeon Lee, and Jong Chul Ye. "Decomposed Diffusion Sampler for Accelerating Large-Scale Inverse Problems". ICLR, 2024.

Thank you for authors' response and clarifications which address several of my concerns. I would keep its original rate.

Reviewer_FB3VOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes UEPS, an unrolled MRI reconstruction model aimed at robustness under domain shift. The author argues that sensitivity map (CSM) is the main bottleneck for gneralization. And to address that, the reconstruction is done in a coil-by-coil manner. And to address the increased processing time, they propose a progressive resolution across cascades, with expanding acquisition matrix size; and a 1D sparse attention pattern algined with the aliasing direction of 1D Cartesian undersampling.

Critical strengths
  1. A relevant problem. Robustness of learned reconstruction to distribution shift is an important practical issue. The observation that learned CSM estimation degrades out of distribution (Table 1, Fig. 4) is useful.
  2. A broad OOD test collection. Ten OOD test sets from five public sources give a reasonably wide evaluation of cross-dataset transfer.
  3. Component ablations. Progressive resolution and sparse attention are each ablated, and the sparse attention pattern is compared against several alternatives.
  4. Good processing speed achievement. Although the 15-coil setting is an underestimate of current common 32-, 64- or even higher channels, especially for brain.
Critical weaknesses
  1. Reconstruction quality is not convincing. Qualitatively, the visual examples show residual artifacts in the UE/UEPS reconstructions that are typical of under-performing reconstructions, scattered all over the image. In comparison, the CSM-based method does show localized deviations, but these are the well-known effects of poorly estimated CSMs. They occur at strong air–tissue interfaces and in low-signal regions (e.g., around the orbits), where CSMs estimated from the low-frequency k-space center are known to be unreliable, and these regions are of limited diagnostic relevance. The spatially scattered residual artifacts of the proposed method are arguably less acceptable for real use than these localized deviations. Quantitatively, the OOD PSNR/SSIM gains alone do not show better image quality. Region-wise error analysis or a reader assessment would be needed. And in distribution performance shall not be ignored here. FastMRI provides public challenge results and learderboards. The in distribution evaluation instead uses the authors’ own protocol (320×320 cropping, volume-max normalization) with retrained baselines, so the numbers cannot be compared with published performance. At least the in distribution comparison should follow the official fastMRI evaluation protocol.
  2. The central premise, that CSM estimation is the primary generalization bottleneck, is not established. The supporting evidence is correlational: Table 1 shows that learned CSMs degrade out of distribution; Fig. 4 shows a few examples where CSM errors co-locate with reconstruction errors; Table 2 shows UE outperforming E2E-VarNet, but removing CSMs also changes the input to the denoiser. No causal test is provided. CSM error could be compared with other shift factors such as contrast, SNR, or coil count. More fundamentally, coil-wise reconstruction does not remove the influence of coil sensitivity. Each coil image is still the object overlayed by its coil profile, so the denoiser must implicitly learn priors over sensitivity-weighted images. This moves the dependence on coil geometry from an explicit, inspectable estimation step into the learned prior. That may make generalization to unseen coil arrays harder, for example when training on 2–28 coils and testing on 32-channel 7T data. The paper should analyze how UE handles unseen coil configurations and how its implicit handling of coil sensitivities compares with explicit, robust CSM estimation.
  3. Limited novelty. Coil-wise (CSM-free) reconstruction. Reconstructing coil images without explicit CSMs is well established in learned MRI reconstruction, e.g., GrappaNet (cited), DeepcomplexMRI (Wang et al., 2020), and scan-specific per-coil k-space networks such as RAKI. The paper’s contribution is therefore limited to studying this design for robustness. It is not compared against these CSM-free approaches beyond GrappaNet in Table 4. Multi-resolution across cascades. The statements that “no prior work has explored multi-resolution across cascades” (Sec. 2.2) and that this is the first design to unify cascaded refinement with resolution scheduling (Sec. 3.3) are not accurate. Coarse-to-fine reconstruction across sequential stages, with data consistency at each scale, has been proposed before. PDAC (Wang et al., 2024), which the paper cites, also decomposes reconstruction progressively over k-space subsampling. Using k-space padding with measured data for upsampling is a reasonable design choice but a modest contribution. Sparse attention. Restricting attention to rows aligned with the aliasing direction is a straightforward instance of axial or factorized attention (e.g., Axial-DeepLab, which is cited).
  4. The more relevant domain shift is not tested. All experiments use 4× equispaced masks with an 8% center, and every input is cropped or padded to 320×320. In practice, the acceleration factor, sampling pattern, and acquisition matrix size together determine which k-space frequencies are sampled. This is arguably a larger and more deployment-relevant shift than anatomy or vendor, and it is exactly what the paper holds fixed. The limitations section acknowledges that unseen masks and acceleration factors are untested, but this undercuts the central robustness claim.
Questions for authors
  1. What causal evidence supports CSM estimation as the primary bottleneck? How much of the OOD gap closes when E2E-VarNet uses ESPIRiT/BART or oracle CSMs?
  2. How does UE generalize to coil arrays whose geometry and channel count differ from training, given that each coil image still carries its sensitivity profile?
  3. How does UEPS perform under unseen acceleration factors (e.g., 8×, 12×), sampling patterns (random, variable-density, Poisson), and acquisition matrix sizes, without retraining?
3 comments
AuthorsOpenReview ↗

Response to Reviewer FB3V Part (1/3)

We thank you for your detailed and constructive feedback. Below, we summarize each concern and respond point by point, grouping related weaknesses (W) and questions (Q).

W1(a): Reconstruction quality is not yet clinically acceptable.

A1: Thanks for emphasizing reconstruction quality and evaluation comparability. We agree with the reviewer that, under these challenging out-of-distribution (OOD) conditions, reconstructions from all methods, including ours, exhibit artifacts on certain test samples, as illustrated in Fig. 5 and Supplementary Figs. 7 and 8. We acknowledge that neither localized nor distributed errors are acceptable for immediate real-world clinical deployment. However, this precisely underscores why rigorous OOD testing and quantitative metrics are crucial: while absolute clinical readiness remains a shared long-term challenge for the field, quantitative comparisons demonstrate that our method makes significant progress and establishes a more robust baseline for this difficult task.


W1(b): In-distribution performance and comparability with the official fastMRI evaluation protocol.

A2: Our common evaluation protocol is intended to enable fair comparisons of reconstruction performance. Published MRI reconstruction studies differ in their data selection, sampling schemes, preprocessing, and metric aggregation, so their reported scores cannot be assumed to be directly comparable. For example, NoVar evaluates fastMRI Brain reconstruction on a T2-only subset restricted to 16-coil acquisitions (Jatyani et al., 2025). Such differences make published scores unsuitable for direct comparison and motivate training and evaluating methods under a common protocol.

To address the concern about ID evaluation, we evaluated UEPS and E2E-VarNet on the fastMRI Brain test set, comprising 558 volumes / 8,852 slices, using the official fastMRI evaluation protocol.

ModelPSNR (dB) ↑SSIM (%) ↑
E2E-VarNet41.712096.2522
UEPS41.622896.1311

For reference, the original E2E-VarNet paper reports 41 dB PSNR and 95.9% SSIM on the fastMRI Brain test set at 4× acceleration (Sriram et al., 2020, Table 3). Note that these published results were obtained using both the fastMRI Brain training and validation splits, whereas our models were trained only on the fastMRI Brain training split. The published values are therefore included as literature references rather than directly matched comparisons.


W2(a) & Q1: Evidence for the role of CSM estimation in OOD degradation and comparisons with BART/ESPIRiT or oracle CSMs.

A3: We trained and evaluated an E2E-VarNet variant using precomputed BART/ESPIRiT maps (Uecker et al., 2015) throughout both training and testing. We report both RSS and MVUE fixed-CSM variants, as MVUE is a more appropriate target for the fixed-CSM setting. The RSS variant uses fully sampled RSS images as both training targets and evaluation references. The MVUE variant uses CSM-based output combination and fully sampled MVUE images as both training targets and evaluation references. All models use identical model configurations and training protocols, differing only in CSM-related settings. The average OOD results are:

ModelOOD PSNR (dB) ↑OOD SSIM (%) ↑
E2E-VarNet (RSS)36.2491.43
E2E-VarNet with fixed BART CSMs (RSS)35.8890.77
E2E-VarNet with fixed BART CSMs (MVUE)36.4291.56
UE-U-Net (RSS)37.1092.48

The RSS-reference comparisons show that UE achieves substantially better OOD reconstruction performance. In the fixed-CSM variant, CSM estimation is external and is not trained jointly with the image-refinement modules. Under domain shift, variations in these externally estimated maps are difficult for the refinement modules to accommodate. These results support coil-wise reconstruction as an effective model-design choice under domain shifts.

AuthorsOpenReview ↗

Response to Reviewer FB3V Part (2/3)

W2(b) & Q2: Model robustness and generalization to unseen coil geometries and channel counts.

A4: UE’s coil-wise denoiser learns a common Fourier deconvolution/de-aliasing task, while CSMs are not involved in defining this convolutional degradation.

Specifically, let xcrefx_c^{\mathrm{ref}} denote the target image for coil cc. Its zero-filled undersampled observation xcx_c is

xc=F−1(MFxcref)+ϵ~c=hM∗xcref+ϵ~c,x_c = \mathcal{F}^{-1}\left(M\mathcal{F}x_c^{\mathrm{ref}}\right) + \tilde{\epsilon}_c = h_M \ast x_c^{\mathrm{ref}} + \tilde{\epsilon}_c,

where MM is the binary sampling mask, F\mathcal{F} is the Fourier transform, hMh_M is the point-spread function determined solely by MM, ∗\ast denotes convolution, and ϵ~c\tilde{\epsilon}_c is image-space noise.

UE directly refines each coil image through

x~c=xc+Dθ(xc),\tilde{x}_{c} = x_{c} + \mathcal{D}_{\theta}(x_{c}),

where Dθ\mathcal{D}_\theta predicts a residual correction directly from the individual coil image xcx_c, and x~c\tilde{x}_c is the refined image. The denoiser's restoration task is to recover xcrefx_c^{\mathrm{ref}} from hM∗xcref+ϵ~ch_M \ast x_c^{\mathrm{ref}} + \tilde{\epsilon}_c through deconvolution and denoising.

By comparison, E2E-VarNet first combines the undersampled coil images using estimated CSMs before passing them to the denoiser. Its refinement step is

x~c=xc+S^c,Dθ(∑jS^j∗xj),\tilde{x}_{c} = x_{c} + \hat{S}_{c}\\,\mathcal{D}_{\theta}\left(\sum_{j} \hat{S}_{j}^{\ast} x_{j}\right),

where S^j\hat{S}_j is the estimated CSM for coil jj. Spatially varying CSM weighting generally does not commute with convolution. The denoiser therefore faces a more complex, sensitivity-dependent effective degradation, rather than the common convolution model encountered in UE. Changes in coil profiles or map-estimation errors can therefore alter both the aliasing structure presented to the denoiser and the updates applied to each coil. These additional map-dependent transformations introduce variation across coil configurations and can make the learned mapping more difficult to generalize. In contrast, UE avoids these transformations through coil-wise reconstruction.

Furthermore, our evaluation already covers unseen coil configurations. For example, the fastMRI Brain training source contains 2–28 coils from Siemens acquisitions at 1.5T/3T, whereas M4Raw GRE uses 4 coils at 0.3T, Stanford 2D covers multiple anatomies with 3–32 coils from GE acquisitions at 3T, and AHEAD uses 32 coils at 7T from Philips. The corresponding RSS-reference PSNR results from the benchmark are:

DatasetCoilsE2E-VarNet PSNR (dB)UEPS PSNR (dB)Gain (dB)
M4Raw GRE435.6237.47+1.85
AHEAD, mean over five echoes3235.3338.46+3.13
Stanford2D3–3238.5440.52+1.98

UEPS achieves the highest reported PSNR on all five AHEAD echoes, exceeding E2E-VarNet by 2.90–3.53 dB. These results provide empirical evidence that UEPS, which uses the UE coil-wise design, generalizes robustly to the unseen acquisition configurations.


W3: Novelty relative to existing CSM-free reconstruction, progressive multi-resolution reconstruction, and axial/factorized attention.

A5: While CSM-free reconstruction has established precedents, to the best of our knowledge, we are the first to systematically investigate independent coil-wise reconstruction for robustness under cross-dataset shifts.

Progressive resolution and sparse attention primarily target efficiency. PDAC decomposes the recovery of missing k-space lines into stages, but each cascade still processes a full-resolution spatial grid. This decomposition therefore does not reduce the per-cascade computational cost (Wang et al., 2024). Our progressive-resolution design instead schedules the actual spatial grid sizes across cascades: early cascades operate on smaller grids, followed by measurement-informed k-space expansion to full resolution. This directly reduces computation in the early cascades. Sparse attention further reduces cost by concentrating computation along the aliasing direction. Together, these designs substantially reduce the inference cost of independent coil-wise reconstruction while maintaining competitive reconstruction quality.

AuthorsOpenReview ↗

Response to Reviewer FB3V Part (3/3)

W4 & Q3: Generalization to unseen acceleration factors, sampling patterns, and acquisition matrix sizes.

A6: Achieving real-world clinical readiness requires robust generalization across a wide spectrum of factors, including hardware configurations, patient populations, anatomies, contrasts, and undersampling patterns. In this work, we push the boundaries by systematically addressing and validating subject- and scanner-level shifts—covering hardware, population, anatomy, and contrast. Generalization to unseen acceleration factors, sampling patterns, and acquisition matrix sizes remains an important avenue for future research.


References

Jatyani, Armeet Singh, et al. "A Unified Model for Compressed Sensing MRI Across Undersampling Patterns". CVPR, 2025.

Sriram, Anuroop, et al. "End-to-End Variational Networks for Accelerated MRI Reconstruction". MICCAI, 2020.

Uecker, Martin, et al. "Berkeley Advanced Reconstruction Toolbox". ISMRM, 2015.

Wang, Chong, et al. "Progressive Divide-and-Conquer via Subsampling Decomposition for Accelerated MRI". CVPR, 2024.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.