- 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
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:
| Dataset | Coils | E2E-VarNet PSNR (dB) | UEPS PSNR (dB) | Gain (dB) |
|---|---|---|---|---|
| M4Raw GRE | 4 | 35.62 | 37.47 | +1.85 |
| AHEAD, mean over five echoes | 32 | 35.33 | 38.46 | +3.13 |
| Stanford2D | 3–32 | 38.54 | 40.52 | +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.
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).
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:
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.
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:
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.
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.
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.