{"id":"eb8fbd66-a9e8-48e3-ac12-f62be11b4b0c","arxiv_id":"2508.13026","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"HierAdaptMR uses hierarchical adapters (protocol-level, center-level, universal) on a variational unrolling backbone to improve cross-center cardiac MRI reconstruction, as evaluated on the CMRxRecon2025 dataset.","lead":"This paper proposes a deep learning method that adapts an MRI reconstruction model to different hospitals, scanners, and imaging protocols using small plug-in modules called adapters. It targets the practical problem of making cardiac MRI reconstruction work on hospitals the model has never seen.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central generalization claim rests on unverified data split and absence of reported results","rationale":"The reader's weakest_assumption correctly flags the unverified data split protocol and the distribution-coverage assumption behind the Universal Adapter. I agree that the data split is the most load-bearing condition, but I would emphasize more strongly that the manuscript text as provided contains no evaluation section at all—so the central claim is not just assumed to be clean, it is entirely unsupported by the available evidence. This is consistent with the reader's UNVERDICTED verdict: there is insufficient information to accept or reject the claim. A concrete, feasible test is to inspect the full paper and challenge split to confirm that 'unseen' means truly held out, and to reproduce the reported comparison against a non-adaptive baseline. No internal inconsistency is apparent from the abstract and introduction, so the verdict should remain unchanged pending that evidence.","tokens_in":1106,"tokens_out":2633,"duration_ms":30696,"concrete_test":"Obtain the full paper and the CMRxRecon2025 challenge specification. Verify that the 5+ test centers are disjoint from all training and validation data, that no test-center information influenced hyperparameter selection or early stopping, and that the reported unseen-center metrics (e.g., PSNR/SSIM) beat a non-adaptive baseline (e.g., the same variational backbone without adapters). If the split is clean and the improvement holds, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's claim of 'superior cross-center generalization' is an empirical claim, but the available manuscript text (arXiv:2508.13026v1) contains only the abstract and introduction. No evaluation section, no quantitative results, no baselines, and no ablation studies are present. The single most load-bearing condition for the central claim is that the 'unseen centers' in CMRxRecon2025 were truly excluded from all model development—including hyperparameter selection, early stopping, and adapter design choices. The abstract's 'stochastic training that learns center-invariant adaptations' describes a plausible domain-randomization strategy, but it does not rule out subtle forms of leakage, such as using challenge leaderboard feedback or tuning hyperparameters against unseen-center validation metrics. Without the data split protocol and the numerical comparisons, the claim of generalization to entirely unseen centers is unverifiable. This is a missing-evidence concern, not an internal inconsistency.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":1333,"tokens_out":1904,"duration_ms":24079,"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":[{"comment":"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":"Abstract"},{"comment":"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":"Section 1 (Introduction)"},{"comment":"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.","section":"Section 1 (Universal Adapter)"}],"minor_comments":[{"comment":"Typo: 'revealssignificant' should be 'reveals significant'.","section":"Section 1"},{"comment":"References [10], [13], [12], [9], and [5] are cited but no bibliography is included in the provided text. A complete reference list is needed.","section":"Section 1"},{"comment":"The claim '5+ centers, 10+ scanners, and 9 modalities' is vague; exact numbers should be reported for reproducibility.","section":"Abstract"},{"comment":"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.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"This submission is at present an extended abstract rather than a full paper. The proposed method is architecturally sensible and the problem is relevant, but the central empirical claim is entirely unverified. I would not reject outright because the missing evidence could be supplied in a revision. The editor may want to consider whether the venue expects complete experimental papers; if so, the lack of any results section is a desk-reject-level issue. I also note that the paper cites works without a bibliography, which suggests an incomplete manuscript rather than a deliberate omission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nWhat 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.\n\nThe 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.\n\nMinor: the abstract says '9 modalities' which seems high for cardiac MRI; possibly a typo or overstatement, but not central.\n\nWho 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.\n\nRecommendation: 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.","headline":"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.","tokens_in":1726,"tokens_out":2425,"would_cite":false,"duration_ms":27823,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Hierarchical adapters let cardiac MRI reconstruction generalize to unseen clinical centers.","keywords":["Cardiac MRI","MRI reconstruction","Domain adaptation","Domain shift","Parameter-efficient adapters","Cross-center generalization","Variational unrolling","Medical imaging"],"falsifier":"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","tokens_in":1053,"feed_emoji":"🧲","tokens_out":4023,"duration_ms":39097,"temperature":0.7,"pith_summary":"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.","feed_headline":"Cardiac MRI model adapts to unseen centers without retraining","feed_subtitle":"Hierarchical adapters split scanner, protocol, and universal features to beat domain shift.","key_machinery":"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.","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Documents the domain-shift problem in cardiac MRI reconstruction that motivates the hierarchical adaptation approach.","marker":"[10]"},{"why":"Provides the CMRxRecon2025 challenge dataset with multiple centers and scanners used for evaluation.","marker":"[13]"},{"why":"Companion challenge description (CMRxRecon2025) that defines the unseen-center evaluation protocol.","marker":"[12]"},{"why":"Describes distinct reconstruction requirements across cardiac MRI sequences, motivating protocol-level adapters.","marker":"[9]"},{"why":"Supplies the reconstruction requirements for perfusion and other protocols that the protocol-level adapters address.","marker":"[5]"}],"fun_headline_variants":["Three adapters let cardiac MRI tackle unseen centers","Hierarchical adapters beat domain shift in cardiac MRI","Cardiac MRI model generalizes to new centers with three adapters","Universal adapter in cardiac MRI handles unseen centers"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Three adapters let cardiac MRI tackle unseen centers","Hierarchical adapters beat domain shift in cardiac MRI","Cardiac MRI model generalizes to new centers with three adapters","Universal adapter in cardiac MRI handles unseen centers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000566,"raw_usage":{"total_tokens":2468,"prompt_tokens":643,"completion_tokens":1825,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":387,"completion_tokens_details":{"reasoning_tokens":1762}},"tokens_in":387,"tokens_out":1825,"duration_ms":14318,"temperature":1.0,"reasoning_tokens":1762,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T19:08:49.534856+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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","supporting_citations":[{"cited_title":"& Others CMRxRecon: A publicly available k-space dataset and benchmark to advance deep learning for cardiac MRI","cited_arxiv_id":null,"evidence_quote":"Documents the domain-shift problem in cardiac MRI reconstruction that motivates the hierarchical adaptation approach."},{"cited_title":"& Others Cmrxrecon2024: A multimodality, multiview k-space dataset boosting universal machine learning for accelerated cardiac mri","cited_arxiv_id":null,"evidence_quote":"Companion challenge description (CMRxRecon2025) that defines the unseen-center evaluation protocol."},{"cited_title":"& Others Black blood myocardial T 2 mapping","cited_arxiv_id":null,"evidence_quote":"Describes distinct reconstruction requirements across cardiac MRI sequences, motivating protocol-level adapters."},{"cited_title":"& Qin, J","cited_arxiv_id":null,"evidence_quote":"Supplies the reconstruction requirements for perfusion and other protocols that the protocol-level adapters address."}],"review_version":1}