HierAdaptMR: Cross-Center Cardiac MRI Reconstruction with Hierarchical Feature Adapters

Ilkay Oksuz; Ruru Xu

REVIEW 3 major objections 4 minor 19 references

HierAdaptMR: Cross-Center Cardiac MRI Reconstruction with Hierarchical Feature Adapters

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Hierarchical adapters let cardiac MRI reconstruction generalize to unseen clinical centers.

desk verdict A sensible hierarchical adapter idea for cross-center cardiac MRI reconstruction, but the arXiv version has no results and the generalization claim is currently unverifiable. read the letter →

arxiv 2508.13026 v1 pith:34JJT2RY submitted 2025-08-18 cs.CV

Ruru Xu , Ilkay Oksuz This is my paper · ORCID

classification cs.CV

keywords CardiacMRIreconstructionDomainadaptationshiftParameter-efficientadaptersCross-centergeneralizationVariationalunrollingMedicalimaging

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tackles the domain-shift problem in deep-learning cardiac MRI reconstruction, where models trained at one hospital fail at others because scanners, protocols, and acceleration patterns differ. It proposes HierAdaptMR, a framework that inserts small parameter-efficient adapters at two levels---protocol-specific and center-specific---on top of a variational unrolling backbone, plus a Universal Adapter trained with stochastic sampling of center and protocol identities. The claim is that this combination learns center-invariant features that transfer to entirely unseen centers without retraining. On the CMRxRecon2025 multi-center benchmark, the authors report superior cross-center generalization while preserving reconstruction quality. A sympathetic reader would take this as evidence that modular adapter-based adaptation is a viable route to deployable cardiac MRI reconstruction.

What carries the argument

The key mechanism is the three-tier adapter stack: Protocol-Level Adapters capture sequence-specific characteristics (cine, LGE, T1/T2 mapping, perfusion), Center-Level Adapters capture scanner-dependent variations, and a Universal Adapter, trained with stochastic sampling of center and protocol identities, learns center-invariant feature transformations. These adapters are parameter-efficient modules inserted into a variational unrolling backbone, so the method adds few parameters while adapting features at multiple granularities. The multi-scale SSIM loss with frequency enhancement and contrast-adaptive weighting stabilises the training.

What would settle it

Retrain HierAdaptMR on a subset of CMRxRecon2025 centers and test it on a center whose scanner vendor, field strength, or sampling pattern does not appear in the training set (e.g., a Canon scanner if training has only GE, Philips, Siemens, UIH). If reconstruction quality on that center degrades significantly relative to centers seen in training, the universal-adapter generalization claim fails. A cheaper check is to partition the data by center, train only on the first half, and measure per-center SSIM/PSNR on the held-out centers while monitoring whether any training-time hyperparameter was

Watch

Extended reading notes

Core claim

HierAdaptMR is a hierarchical feature adaptation framework for cardiac MRI reconstruction under multi-level domain shift. It decomposes domain variation into protocol-level (sequence-specific) and center-level (scanner-dependent) components, each handled by lightweight adapters inserted into a variational unrolling reconstruction network. A third Universal Adapter is trained by stochastically sampling center and protocol identities, forcing it to learn transformations that are invariant to these identities. Combined with a multi-scale SSIM loss, frequency-domain enhancement, and contrast-adaptive weighting, the framework is claimed to achieve strong generalization to centers unseen during tr

Load-bearing premise

The claim of generalization to unseen centers rests on the assumption that the stochastic sampling of center and protocol identities during training produces features that cover the distribution of any future center, and that the reported 'unseen' centers were never used in any form of model or hyperparameter selection.

Editorial extensions

If this is right

If the claim holds, cardiac MRI reconstruction models can be deployed to a new hospital without retraining on local data, provided the new center's acquisition settings fall within the learned domain variation.
Protocol-level adapters allow one model to serve multiple imaging sequences, removing the need for sequence-specific models.
The parameter-efficient design means the adaptation adds only a small number of trainable parameters, making it feasible to update adapters on-site.
The stochastic training of the Universal Adapter suggests a general recipe for learning domain-invariant representations in other medical imaging tasks.
The multi-scale SSIM and frequency-domain losses indicate that reconstruction quality is maintained not just in pixel space but perceptually.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The hierarchical decomposition implies that if a new center introduces a previously unseen scanner vendor, only a new Center-Level Adapter would need to be trained, not the whole model; the paper does not explicitly test this.
The stochastic sampling of center/protocol identities is a form of domain randomization; a natural extension is to test the Universal Adapter's sensitivity to the diversity of the training centers (e.g., how many centers are needed before transfer stabilises).
The paper's evaluation on CMRxRecon2025 cannot be independently verified from the abstract alone; a critical test would be a leave-one-center-out analysis with explicit reporting of which centers were held out and whether any hyperparameters were tuned on them.
The contrast-adaptive weighting may also help with variable acceleration factors (8x-24x), suggesting a potential extension to adaptive acceleration-aware reconstruction.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is an architecture proposal with no experimental section in the version I saw. The hierarchical adapter idea is sensible and the problem is real, but the abstract's 'superior cross-center generalization' is a claim, not a demonstrated result.

What is actually new: the three-level decomposition into protocol, center, and universal adapters on a variational unrolling backbone is a specific combination I don't recall from the CMRxRecon challenge papers. The universal adapter trained with stochastic sampling of center/protocol identities is a neat domain-randomization trick, and parameter-efficient adapters are practical for deployment. The intro also does a good job laying out the levels of domain shift (scanner, protocol, acquisition). Credit where it's due: the authors clearly see the right problem.

The soft spots are mostly about evidence. There are no numbers, no baselines, no ablations, no described data split. You can't check whether the 'unseen centers' were truly held out from hyperparameter tuning or early stopping. The stress-test note about leakage is legitimate but unanswerable from the text. That's a missing-evidence problem, not an internal contradiction. The modeling assumption that stochastic training yields center-invariant features is plausible but unproven; the paper would be stronger with an ablation showing the universal adapter alone vs. the full hierarchy.

Minor: the abstract says '9 modalities' which seems high for cardiac MRI; possibly a typo or overstatement, but not central.

Who is this for: people working on cross-center medical imaging reconstruction, especially challenge participants. As an arXiv v1, it's a position statement. If there's a full version with experiments (challenge paper or later revision), it deserves a serious referee. As it stands, there isn't enough content to peer review.

Recommendation: ask the authors to post the full version with results before sending it to review. The idea is worth engaging with, but not the paper in this form.

Referee Report

3 major / 4 minor

Summary. The manuscript proposes HierAdaptMR, a hierarchical feature adaptation framework for cardiac MRI reconstruction under domain shift. It combines Protocol-Level Adapters for sequence-specific characteristics, Center-Level Adapters for scanner-dependent variations, and a Universal Adapter trained with stochastic sampling to generalize to unseen centers. The backbone is a variational unrolling network, trained with a multi-scale SSIM loss with frequency-domain enhancement and contrast-adaptive weighting. The central claim is that evaluation on the CMRxRecon2025 dataset across 5+ centers, 10+ scanners, and 9 modalities demonstrates superior cross-center generalization while maintaining reconstruction quality. The submitted text, however, contains only the abstract and introduction; no experimental methods, results, baselines, ablations, or data-split details are provided.

Significance. If the claimed results hold, the hierarchical adapter design is a plausible and potentially useful contribution to cross-center cardiac MRI reconstruction, a clinically relevant problem. The proposed separation of protocol-level and center-level adapters is architecturally reasonable, and the use of a universal adapter with stochastic training is a sensible domain-randomization strategy. The code URL is a positive step toward reproducibility. However, the manuscript as submitted does not include the evidence needed to assess the central empirical claim. No quantitative results, error bars, baselines, ablations, or evaluation-protocol details are present, so the significance cannot currently be evaluated beyond the plausibility of the architecture.

major comments (3)

[Abstract] The central claim, 'Comprehensive evaluation on the CMRxRecon2025 dataset ... demonstrates superior cross-center generalization while maintaining reconstruction quality,' is an empirical assertion, but the manuscript contains no evaluation section, no quantitative results, no baselines, and no ablations. This is the load-bearing claim of the paper, and it is currently unsupported by any verifiable evidence. A full experimental section with comparisons to relevant single-center and cross-center baselines, per-center and per-sequence results, and statistical error bars must be added.
[Section 1 (Introduction)] The paper states that CMRxRecon2025 evaluates models on 'entirely unseen centers,' but it does not describe the data-split protocol or the role of the challenge validation set in model development. For the cross-center generalization claim to be meaningful, the manuscript must establish that center identity, scanner parameters, and protocol information of the unseen centers were not used for any design decision—including adapter architecture choices, bottleneck dimensions, loss weights, stochastic training probabilities, or early stopping. If hyperparameters were tuned on challenge validation metrics that include unseen-center data, the generalization claim would be substantially weakened. The authors should state explicitly how the training, validation, and evaluation centers were partitioned.
[Section 1 (Universal Adapter)] The Universal Adapter relies on stochastic training over center and protocol identities to learn 'center-invariant adaptations.' This assumes that the training set's domain distribution adequately covers or represents the deployment distribution of unseen centers. The manuscript provides no analysis of how many centers/scanners/protocols were used for training versus evaluation, nor any discussion of how domain coverage was ensured. A concrete test of this assumption would be an ablation varying the number of training centers and reporting generalization performance, or per-center results to show that the universal adapter does not simply memorize one or two dominant domains.

minor comments (4)

[Section 1] Typo: 'revealssignificant' should be 'reveals significant'.
[Section 1] References [10], [13], [12], [9], and [5] are cited but no bibliography is included in the provided text. A complete reference list is needed.
[Abstract] The claim '5+ centers, 10+ scanners, and 9 modalities' is vague; exact numbers should be reported for reproducibility.
[General] The manuscript is only an abstract and an introduction. Even for a short paper, the structure should include a methods section describing the adapter architectures and the loss function in detail, or explicitly state that the full paper is forthcoming.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the manuscript contains only an abstract and introduction with an empirical method claim; there is no fitted-then-predicted quantity, no uniqueness theorem, and no load-bearing self-citation chain.

full rationale

The submitted text (arXiv:2508.13026v1) consists of the abstract and Section 1 (Introduction). There are no equations, no parameter-fitting steps, no reported numerical results, and no derivation of a prediction from a fit. The central claim—'Comprehensive evaluation on the CMRxRecon2025 dataset ... demonstrates superior cross-center generalization'—is an empirical performance assertion, not a derivation. The method description ('Protocol-Level Adapters', 'Center-Level Adapters', 'Universal Adapter ... through stochastic training') is a stated design, and no argument is made that would make a result follow from its own definition. The single most load-bearing condition for the empirical claim, that CMRxRecon2025's unseen centers were truly excluded from all model development, is not verifiable from the available manuscript; however, missing evidence about the data split is a correctness and reporting concern, not circularity. No self-citations are used as load-bearing premises; citations in the visible text ([10], [13], [12], [9], [5]) are external context for domain shift and the challenge. Since no circular step can be quoted or exhibited, the appropriate finding is no significant circularity (score 0).

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim is empirical, so no first-principles derivation exists. The listed free parameters are hyperparameters that must be tuned, and the axioms are the domain assumptions about dataset representativeness and honest data splits. No new physical or mathematical entities are introduced; the adapters are architectural modules.

free parameters (3)

Adapter bottleneck dimensions
Not disclosed in abstract; typical hyperparameters controlling capacity of the adapters.
Loss weighting coefficients
Multi-scale SSIM, frequency domain, and contrast-adaptive weighting imply tuned weights; values not reported.
Stochastic training sampling probabilities
Universal Adapter's stochastic training requires sampling probabilities for centers/protocols; not specified.

assumptions (4)

standard math Backpropagation and standard deep learning optimization converge to a useful solution
All deep learning papers rely on gradient-based optimization without proof of global optimality.
domain assumption CMRxRecon2025 dataset is representative of deployment conditions
Abstract claims 5+ centers, 10+ scanners, 9 modalities; assumes these cover unseen center variability.
domain assumption The official held-out centers are truly unseen and not used in any model selection
The claim of generalization to unseen centers requires that the split is honest; not verifiable from abstract.
domain assumption Adapters are sufficient to capture domain shift without retraining the backbone
The method's premise is that parameter-efficient adapters at protocol and center levels can encode domain-specific features while keeping shared backbone.

how reviews work

0 comments

Cite this review

Pith. "Pith review of HierAdaptMR: Cross-Center Cardiac MRI Reconstruction with Hierarchical Feature Adapters." pith.science (2026). https://pith.science/paper/34JJT2RY

@misc{pith2026250813026,
  author       = {Pith},
  title        = {Pith review of: HierAdaptMR: Cross-Center Cardiac MRI Reconstruction with Hierarchical Feature Adapters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/34JJT2RY}},
  note         = {Machine review of arXiv:2508.13026}
}

read the original abstract

Deep learning-based cardiac MRI reconstruction faces significant domain shift challenges when deployed across multiple clinical centers with heterogeneous scanner configurations and imaging protocols. We propose HierAdaptMR, a hierarchical feature adaptation framework that addresses multi-level domain variations through parameter-efficient adapters. Our method employs Protocol-Level Adapters for sequence-specific characteristics and Center-Level Adapters for scanner-dependent variations, built upon a variational unrolling backbone. A Universal Adapter enables generalization to entirely unseen centers through stochastic training that learns center-invariant adaptations. The framework utilizes multi-scale SSIM loss with frequency domain enhancement and contrast-adaptive weighting for robust optimization. Comprehensive evaluation on the CMRxRecon2025 dataset spanning 5+ centers, 10+ scanners, and 9 modalities demonstrates superior cross-center generalization while maintaining reconstruction quality. code: https://github.com/Ruru-Xu/HierAdaptMR

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 16 canonical work pages

[1]

& Ebrahimi Moghaddam, M

Anvari Hamedani, K., Razizadeh, N., Nabavi, S. & Ebrahimi Moghaddam, M. An All-in-One Approach for Accelerated Cardiac MRI Reconstruction. International Workshop On Statistical Atlases And Computational Models Of The Heart . pp. 464-475 (2024)

work page 2024
[2]

Statistical Distance-Guided Unsupervised Domain Adaptation for Automated Multi-Class Cardiovascular Magnetic Resonance Image Quality Assessment

Nabavi, S., Hamedani, K., Moghaddam, M., Abin, A. & Frangi, A. Statistical distance-guided unsupervised domain adaptation for automated multi-class cardiovascular magnetic resonance image quality assessment. ArXiv Preprint ArXiv:2409.00375 . (2024)

work page Pith review arXiv 2024
[3]

& Wang, C

Lyu, J., Tong, X. & Wang, C. Parallel imaging with a combination of SENSE and generative adversarial networks (GAN). Quantitative Imaging In Medicine And Surgery . 10, 2260-2273 (2020)

work page 2020
[4]

& Qin, J

Lyu, J., Tian, Y., Cai, Q., Wang, C. & Qin, J. Adaptive channel-modulated personalized federated learning for magnetic resonance image reconstruction. Computers In Biology And Medicine . 165 pp. 107330 (2023)

work page 2023
[5]

& Qin, J

Lyu, J., Wang, S., Tian, Y., Zou, J., Dong, S., Wang, C., Aviles-Rivero, A. & Qin, J. STADNet: Spatial-temporal attention-guided dual-path network for cardiac cine MRI super-resolution. Medical Image Analysis . 94 pp. 103142 (2024)

work page 2024
[6]

& Pezzotti, N

Lyu, D., Rao, C., Staring, M., Osch, M., Doneva, M., Lamb, H. & Pezzotti, N. UPCMR: A Universal Prompt-Guided Model for Random Sampling Cardiac MRI Reconstruction. International Workshop On Statistical Atlases And Computational Models Of The Heart . pp. 453-463 (2024)

work page 2024
[7]

& Rueckert, D

Qin, C., Schlemper, J., Caballero, J., Price, A., Hajnal, J. & Rueckert, D. Convolutional recurrent neural networks for dynamic MR image reconstruction. IEEE Transactions On Medical Imaging . 38, 280-290 (2018)

work page 2018
[8]

& Rueckert, D

Qin, C., Duan, J., Hammernik, K., Schlemper, J., Küstner, T., Botnar, R., Prieto, C., Price, A., Hajnal, J. & Rueckert, D. Complementary time-frequency domain networks for dynamic parallel MR image reconstruction. Magnetic Resonance In Medicine . 86, 3274-3291 (2021)

work page 2021

Show all 19 references

[9]

& Others Black blood myocardial T 2 mapping

Wang, C., Jang, J., Neisius, U., Nezafat, M., Fahmy, A., Kang, J., Rodriguez, J., Goddu, B., Pierce, P., Berg, S. & Others Black blood myocardial T 2 mapping. Magnetic Resonance In Medicine . 81, 153-166 (2019)

2019
[10]

& Others CMRxRecon: A publicly available k-space dataset and benchmark to advance deep learning for cardiac MRI

Wang, C., Lyu, J., Wang, S., Qin, C., Guo, K., Zhang, X., Yu, X., Li, Y., Wang, F., Jin, J. & Others CMRxRecon: A publicly available k-space dataset and benchmark to advance deep learning for cardiac MRI. Scientific Data . 11, 687 (2024)

2024
[11]

& Wen, B

Wang, C., Guo, L., Wang, Y., Cheng, H., Yu, Y. & Wen, B. Progressive divide-and-conquer via subsampling decomposition for accelerated mri. Proceedings Of The IEEE/CVF Conference On Computer Vision And Pattern Recognition . pp. 25128-25137 (2024)

2024
[12]

& Others Cmrxrecon2024: A multimodality, multiview k-space dataset boosting universal machine learning for accelerated cardiac mri

Wang, Z., Wang, F., Qin, C., Lyu, J., Ouyang, C., Wang, S., Li, Y., Yu, M., Zhang, H., Guo, K. & Others Cmrxrecon2024: A multimodality, multiview k-space dataset boosting universal machine learning for accelerated cardiac mri. Radiology: Artificial Intelligence . 7, e240443 (2025)

2025
[13]

& Others Towards Universal Learning-based Model for Cardiac Image Reconstruction: Summary of the CMRxRecon2024 Challenge

Wang, F., Wang, Z., Li, Y., Lyu, J., Qin, C., Wang, S., Guo, K., Sun, M., Huang, M., Zhang, H. & Others Towards Universal Learning-based Model for Cardiac Image Reconstruction: Summary of the CMRxRecon2024 Challenge. ArXiv Preprint ArXiv:2503.03971 . (2025)

2025
[14]

& Metaxas, D

Xin, B., Ye, M., Axel, L. & Metaxas, D. Fill the k-space and refine the image: Prompting for dynamic and multi-contrast MRI reconstruction. International Workshop On Statistical Atlases And Computational Models Of The Heart . pp. 261-273 (2023)

2023
[15]

& Metaxas, D

Xin, B., Ye, M., Axel, L. & Metaxas, D. Rethinking deep unrolled model for accelerated MRI reconstruction. European Conference On Computer Vision . pp. 164-181 (2024)

2024
[16]

& Oksuz, I

Xu, R., Özer, C. & Oksuz, I. Hypercmr: Enhanced multi-contrast cmr reconstruction with eagle loss. International Workshop On Statistical Atlases And Computational Models Of The Heart . pp. 152-163 (2024)

2024
[17]

& Oksuz, I

Xu, R. & Oksuz, I. A reinforcement learning approach for optimized MRI sampling with region-specific fidelity. Neurocomputing . 638 pp. 130116 (2025)

2025
[18]

, " * write output.state after.block = add.period write

ENTRY address author booktitle chapter doi edition editor eid howpublished institution journal key month note number organization pages publisher school series title type url volume year label INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION in...
[19]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

& Ebrahimi Moghaddam, M

Anvari Hamedani, K., Razizadeh, N., Nabavi, S. & Ebrahimi Moghaddam, M. An All-in-One Approach for Accelerated Cardiac MRI Reconstruction. International Workshop On Statistical Atlases And Computational Models Of The Heart . pp. 464-475 (2024)

work page 2024

[2] [2]

Statistical Distance-Guided Unsupervised Domain Adaptation for Automated Multi-Class Cardiovascular Magnetic Resonance Image Quality Assessment

Nabavi, S., Hamedani, K., Moghaddam, M., Abin, A. & Frangi, A. Statistical distance-guided unsupervised domain adaptation for automated multi-class cardiovascular magnetic resonance image quality assessment. ArXiv Preprint ArXiv:2409.00375 . (2024)

work page Pith review arXiv 2024

[3] [3]

& Wang, C

Lyu, J., Tong, X. & Wang, C. Parallel imaging with a combination of SENSE and generative adversarial networks (GAN). Quantitative Imaging In Medicine And Surgery . 10, 2260-2273 (2020)

work page 2020

[4] [4]

& Qin, J

Lyu, J., Tian, Y., Cai, Q., Wang, C. & Qin, J. Adaptive channel-modulated personalized federated learning for magnetic resonance image reconstruction. Computers In Biology And Medicine . 165 pp. 107330 (2023)

work page 2023

[5] [5]

& Qin, J

Lyu, J., Wang, S., Tian, Y., Zou, J., Dong, S., Wang, C., Aviles-Rivero, A. & Qin, J. STADNet: Spatial-temporal attention-guided dual-path network for cardiac cine MRI super-resolution. Medical Image Analysis . 94 pp. 103142 (2024)

work page 2024

[6] [6]

& Pezzotti, N

Lyu, D., Rao, C., Staring, M., Osch, M., Doneva, M., Lamb, H. & Pezzotti, N. UPCMR: A Universal Prompt-Guided Model for Random Sampling Cardiac MRI Reconstruction. International Workshop On Statistical Atlases And Computational Models Of The Heart . pp. 453-463 (2024)

work page 2024

[7] [7]

& Rueckert, D

Qin, C., Schlemper, J., Caballero, J., Price, A., Hajnal, J. & Rueckert, D. Convolutional recurrent neural networks for dynamic MR image reconstruction. IEEE Transactions On Medical Imaging . 38, 280-290 (2018)

work page 2018

[8] [8]

& Rueckert, D

Qin, C., Duan, J., Hammernik, K., Schlemper, J., Küstner, T., Botnar, R., Prieto, C., Price, A., Hajnal, J. & Rueckert, D. Complementary time-frequency domain networks for dynamic parallel MR image reconstruction. Magnetic Resonance In Medicine . 86, 3274-3291 (2021)

work page 2021