PaperlayerView exact revision on arXiv

Source-linked AI summary

Evaluating the Safety of Deep Learning-Based Brain MRI Reconstruction

Dat Tat Mai, Thai Viet Pham, Thu Nguyen Thi Dang, James Jin Kang

arXiv:2608.28714v1eess.IVcs.CV

TL;DR

Pixel-wise metrics and diagnostic assessment are seldom compared directly in deep learning brain MRI reconstruction, leaving a central evaluation question unresolved. This review synthesizes evaluation practices across 263 studies and finds that current paradigms do not settle the safety question. The authors identify unaudited extraction as a key limitation.

  • Problem

    A systematic account of how often reconstruction studies measure metric–reader agreement, why metrics may fail, and what current evaluation can establish has been missing.

    Source: §II · p. 3

  • Method

    The review searched seven databases without a date limit and used a sensitivity-oriented strategy to identify evidence across clinical and engineering vocabularies.

    Source: §III · p. 3

  • Results

    Only 4.9% of studies did both kinds of evaluation, reflecting a division between communities with access to raw k-space and those with pathology, readers, and acute presentations.

    Source: §V · p. 18

  • Takeaways & Limitations

    No evaluation paradigm currently in use settles the safety question by itself.

    Source: §V · pp. 17–18

    Extraction has no reliability estimate; released supporting quotations make assignments auditable but do not substitute for statistical reliability.

    Source: B. Threats to validity · p. 19

Abstract

from arXiv · show

Objective: Deep learning accelerates brain MRI four- to tenfold, but models can erase lesions or synthesize false tissue - failures pixel-averaged metrics like PSNR and SSIM miss. We review whether current evaluation practices detect this blind spot. Methods: Following PRISMA 2020, we searched seven databases without date limits, including 263 studies (1995-2026), appraised them using QUADAS-2 and matched instruments, and synthesized narratively. Categories were derived from titles, abstracts, and controlled vocabulary; reported prevalence figures represent floors. Duplicate screening achieved high agreement (Fleiss kappa = 0.877), as did appraisal (0.788; 0.390 where observable). Extraction is unaudited. Results: Only 18 of 263 studies (6.8%) recorded both a fidelity metric and reader assessment on identical data, leaving the central surrogate unmeasured. Reader studies mostly measured inter-reader agreement, which was weak: fastMRI 2020 concordance reached 0.457 and 0.386 (Kendall W), improving only where SSIM diverged. Erasing a 100 mm3 lacunar infarct shifts global PSNR by 0.03 dB under the stated error model. As the corpus grew fivefold, reader assessments dropped from 32% to 18%, recovering to 21%. Generative models - most associated with hallucination (39%) - were among the least reader-evaluated (11.3%), while self-supervised models reached 47% with zero reader evaluation. Only 5% released code and ran reader studies; none evaluated a model observer; no named dataset covered acute stroke or hemorrhage. Conclusions: On these floors, current evaluation practices cannot certify diagnostic safety. We derive five requirements safety-oriented evaluations must meet.

I. INTRODUCTION

Deep learning accelerates brain MRI, but plausible reconstructions can erase real findings or invent tissue that averaged image-quality metrics may miss. This review examines the evidence behind that blind spot, the limits of current evaluation approaches, and requirements for safety-oriented assessment.

I. INTRODUCTION

  • Learned priors accelerate MRI reconstruction, but can smooth away faint structure, produce artifacts under motion or scanner shifts, and invent plausible tissue.PSNR and SSIM only partially track radiologist quality judgements, while a small lesion barely shifts a whole-volume average.Source: §I · pp. 1–2
  • The review synthesizes 263 studies, four mathematical mechanisms of metric divergence, the evaluation gap’s structural context, and five requirements for safety-oriented evaluation.It diagnoses what current instruments can detect; it does not estimate how often deployed reconstructions erase lesions, and the proposed auditing framework is addressed in a companion paper.Source: §I · p. 2
  • Prior surveys catalogue architectures and benchmark fidelity, but do not tabulate primary studies’ evaluation practices or synthesize metric–reader agreement.This review asks how reliably conventional metrics track radiologist ratings, why they may miss focal pathology, and what current paradigms establish.Source: §II · p. 2, §I · p. 2

II. RELATED WORK

  • Task-based image-quality assessment ties image quality to performance on a specified diagnostic task, but no included study evaluated a model observer.A task-specific detectability index can respond to focal changes that whole-volume L2 scores average away; specifying the signal and background remains difficult for learned reconstruction.Source: §II · p. 2
  • Work on artifacts, motion, distribution shift, hallucination, and counterfactual testing remains divided across literatures and has not been connected to review-level evidence on reconstruction evaluation.Prior hallucination analyses place prior-induced structure in the forward operator’s null space, where measurement-space consistency checks cannot see it.Source: §II · p. 2
  • The review’s four mechanisms combine elementary metric properties and established results; its contribution is their quantification and assembly against the 263-study corpus.The identity introduced at Equation (2) bounds the PSNR change from erasing a lesion of any given volume.Source: §II · pp. 2–3
  • Available resources do not combine raw k-space with acute vascular pathology, separating engineering benchmarks from clinical reader studies.The review’s novelty is assembling how often metric–reader agreement is measured, why metrics can fail, and what current evaluation paradigms establish.Source: §II · p. 3

III. METHODOLOGY

The review followed PRISMA 2020, searched seven databases without a date limit, and included 263 studies after screening. Reviewers extracted reported evaluation practices and appraised studies using design-matched instruments; evidence was synthesized narratively.

III. METHODOLOGY

  • Seven databases were searched on 23–24 July 2026 without a publication-date limit, using broad concept groups to capture reconstruction, safety, failure modes, and evaluation.The sensitivity-oriented strategy accepted a large screening burden to avoid missing evidence reported in passing or using distinct clinical and engineering vocabularies.Source: §III · p. 3
  • 403 duplicate records were removed from 27,327 identified, leaving 26,924 unique records for screening.The authors note that one preprint–journal duplicate survived matching because its metadata differed.Source: §III · p. 3
  • 263 primary studies met the inclusion criteria after full-text verification; 81 of 1,149 advanced reports could not be obtained.Three reviewers screened against pre-specified PICOS criteria, with flagged or borderline decisions escalated for review.Source: §III · pp. 3–4
  • Fleiss’ κ was 0.877 across 214 duplicate-screened records, while appraisal agreement was κ = 0.788 overall and κ = 0.390 in the second round.The appraisal coefficient averaged two rounds that differed: the first-round subset was unanimous, while the second comprised studies deferred as less straightforward.Source: §III · pp. 5–6
  • Reviewers extracted study characteristics with verbatim supporting quotations, assigning method, failure-mode, and evaluation categories from titles, abstracts, and controlled vocabulary.Unreported variables were recorded as not reported rather than inferred, so prevalence counts are lower bounds.Source: §III · p. 5
  • Study quality was appraised with QUADAS-2 for reader and diagnostic-accuracy studies, descriptive STARD checks, and a reproducibility and robustness appraisal for algorithmic studies.Because protocols, architectures, datasets, and outcome measures were heterogeneous, the review synthesized findings narratively without pooling effects or fitting a meta-analytic model.Source: §III · pp. 5–6

IV. RESULTS

The reported counts describe the studies included and reports assessed, not the full literature, because 81 candidate reports remained unretrievable.

IV. RESULTS

  • 81 candidate reports remained unretrievable, so the 1,068 assessed reports and 263 included studies represent lower bounds on the literature.The authors state that counts and prevalence statements describe those assessed and included records rather than estimates of the full literature.Source: §IV · p. 6
  • 1,068 of 1,149 reports advanced to full-text assessment were obtained and assessed.The remaining 81 could not be retrieved.Source: §IV · p. 6
  • 263 studies were included in the review, with prevalence counts bounded by the unretrieved reports.The reported counts describe the 1,068 reports assessed in full and the included studies.Source: §IV · p. 6

A. Selection and corpus

The review identified 27,327 records across seven databases and included 263 primary studies published from 1995–2026. The included corpus combines QUADAS-2 appraisal of reader and diagnostic studies with reproducibility and robustness appraisal of algorithmic studies.

A. Selection and corpus

  • 27,327 records from seven databases were deduplicated to 26,924, of which 1,149 proceeded to full-text eligibility.Full texts were obtained for 1,068 of the 1,149 reports (93.0%).Source: A. Selection and corpus · p. 6
  • All 57 diagnostic and reader studies were appraised with QUADAS-2, while the other 206 studies received reproducibility and robustness appraisal.Table III pools the first and second appraisal rounds; six of 399 domain judgments with no majority were recorded as Unclear.Source: Table III · p. 7
  • Tables IV and V report QUADAS-2 risk-of-bias and applicability judgments, respectively, for the 57 appraised reader and diagnostic studies.Source: Table IV · p. 7, Table V · p. 8
  • 263 primary studies from 1995–2026 remained after 1,068 reports were assessed and 787 were excluded.The assessed reports comprised 785 PICOS exclusions and two duplicate reports of included studies; 18 additional reports were recorded as background.Source: A. Selection and corpus · p. 8
  • The included corpus grew from 34 studies in 1995–2019 to 174 in 2023–2026, with 55 published in 2020–2022.Source: A. Selection and corpus · p. 8

B. RQ1: the metric–reader comparison is almost never made, and where readers appear they agree only partially

Metric–reader agreement is rarely measured on the same data, and the available comparisons show only partial, condition-dependent agreement. Reader assessments also illustrate that aggregate quality judgments can coexist with less conspicuous focal findings.

B. RQ1

  • The available metric–reader comparisons show partial, condition-dependent agreement, and only two of four Table VII rows used undersampled-k-space reconstructions.The direct evidence for this comparison in the review’s setting therefore rests on two on-target rows.Source: B. RQ1: the metric–reader comparison… · p. 8
  • FastMRI 2020 reader concordance was 0.457 in the 4× track and 0.386 in the 8× track, rising to 0.781 in the transfer track.The organisers reported improved concordance as SSIM scores diverged.Source: B. RQ1: the metric–reader comparison… · p. 8
  • In one of 66 paired examinations, a possible acute-to-subacute infarct was less conspicuous on accelerated DWI despite no overall diagnostic-quality difference between protocols.The lesion was smaller than the section thickness, and the authors attributed the discrepancy to image-quality difference and section positioning; one case is not a prevalence estimate.Source: B. RQ1: the metric–reader comparison… · pp. 8–9
  • Across the corpus, pixel-wise metrics (PSNR/SSIM/NRMSE) were recorded in 84 studies (47.5% of recorded; legacy fidelity reporting), while observer/reader studies numbered 57 (32.2% of recorded; verified subset).Source: B. RQ1: the metric–reader comparison… · p. 9
  • Only 18 of 263 studies (6.8%) reported both a pixel-wise fidelity metric and reader assessment on the same data.This is a screening-level lower bound and the minimum precondition for measuring metric–reader agreement.Source: B. RQ1: the metric–reader comparison… · p. 9
  • No retrieved study reported a model observer, although the review’s search vocabulary could have missed studies described only by observer names.The corpus therefore supports a bounded claim about retrieved and extracted studies, not a definitive field-wide count.Source: B. RQ1: the metric–reader comparison… · p. 9

C. RQ2: four mechanisms explain the divergence

Four mechanisms explain how whole-volume fidelity metrics can miss focal diagnostic errors: reference dependence, spatial averaging, error-sign symmetry, and realism–reward trade-offs. For a resolvable acute infarct, the modeled global PSNR change is near conventional reporting precision.

Spatial averaging

  • Conventional L2 metrics average residual error uniformly across all voxels, diluting a focal lesion error within the whole-volume score.With lesion fraction f and local error ratio r, the global PSNR change follows directly from the stated identity.Source: C. RQ2: four mechanisms explain the… · p. 12
  • A 100 mm3 acute lacunar infarct erased at a 100-fold local error ratio changes global PSNR by 0.027–0.036 dB under the stated error model.That is about three hundredths of a decibel, near the precision of a score conventionally reported to two decimals; the bound is computed, not measured.Source: C. RQ2: four mechanisms explain the… · p. 12

Loss-function symmetry

  • MSE and SSIM treat positive and negative errors symmetrically, although suppressing a true hyperintense lesion can cause a missed finding.The review notes lesion-mask-weighted, overlap-weighted, and gradient-based losses, but none of those studies paired its correction with reader assessment.Source: C. RQ2: four mechanisms explain the… · p. 12

Realism–reward trade-off

  • Adversarial and perceptual objectives reward outputs resembling the training distribution, which may add plausible texture absent from the acquired measurement.This realism–reward trade-off is associated with generative reconstruction.Source: C. RQ2: four mechanisms explain the… · p. 12

Reference dependence

  • All four metrics assume a registered, artifact-free reference; intra-scan motion invalidates that assumption, and 69 studies relied solely on fully sampled acquisitions as their reference standard.That reference is valid only where the subject did not move.Source: C. RQ2: four mechanisms explain the… · p. 12

D. RQ3: what current evaluation paradigms can and cannot establish

Current evaluation paradigms answer questions adjacent to reconstruction safety rather than establishing safety itself. Task-based observer assessment addresses the diagnostic question directly but is scarce, while null-space inspection is absent from the corpus.

What the four empirical paradigms establish

  • Task-based observer assessment directly addresses diagnostic change, but only 57 of 263 studies used it and just 18 paired it with a fidelity metric.It is the only paradigm in the review that settles the safety question directly.Source: D. RQ3: what current evaluation… · p. 12
  • Perceptual and rank-trained metrics track reader preference better than MSE or SSIM as optimization targets, but inherit reader variability and lack demonstrated transfer to reconstruction.The one such study used synthetically corrupted images rather than undersampled-k-space reconstructions.Source: D. RQ3: what current evaluation… · p. 12
  • Uncertainty estimates indicate model uncertainty, not diagnostic correctness; a prior can remove a lesion confidently, leaving low posterior variance where the error occurred.Pixel-wise fidelity metrics likewise establish aggregate reference agreement, not focal change.Source: D. RQ3: what current evaluation… · p. 12

Two limits that follow from the physics

  • A measurement-space residual is unchanged by edits in the acquisition operator's null space, including erased lesions or invented vessels.Undersampling creates that null space, which grows as acceleration becomes more aggressive.Source: D. RQ3: what current evaluation… · pp. 12–13
  • Inspecting the null-space component is a natural way to look for erasure and hallucination, but no corpus study used this paradigm.Computing the projection requires the forward operator, raw k-space, and coil sensitivities, which commercial inline pipelines do not export.Source: D. RQ3: what current evaluation… · p. 13

E. Structure of the evaluation gap

Safety-relevant evaluation is scarce and unevenly distributed across publication eras, method families, and research communities. The named public datasets do not provide a benchmark curated for acute ischemic stroke or intracranial haemorrhage.

Public datasets and acute pathology

  • No study in the corpus benchmarked against a dataset curated for acute ischemic stroke or intracranial haemorrhage.fastMRI+ labels do not distinguish acute from chronic infarction, and its annotated series omit DWI and SWI/T2*-weighted imaging.Source: E. Structure of the evaluation gap · p. 14
  • Stroke-lesion datasets such as ISLES and ATLAS are segmentation benchmarks distributed as processed images without raw k-space, limiting their use for reconstruction evaluation.The dataset discussion distinguishes their existence from their suitability as reconstruction benchmarks.Source: E. Structure of the evaluation gap · p. 14

Rigour is divided between two literatures

  • Only 13 of 263 studies (4.9%) released code and reported reader assessment; 122 (46.4%) did neither.Clinical imaging venues more often reported reader assessment, while engineering and methods venues more often released code; the association is descriptive because venue groups were classified post hoc.Source: E. Structure of the evaluation gap · pp. 14–15

Evaluation across eras and method families

  • Generative studies recorded hallucination or fidelity instability in 31/80 (38.8%) but reader assessment in only 9/80 (11.3%); self-supervised studies recorded 47.1% hallucination and no reader assessment.These are recorded-label lower bounds; the self-supervised group contained 17 studies.Source: E. Structure of the evaluation gap · p. 15, Table XI · p. 15
  • Reader-assessment share fell from 32.4% in 1995–2019 to 18.2% in 2020–2022, then recovered only to 20.7% in 2023–2026 as the corpus grew fivefold.In the latest era, reader assessment was 52.0% (13 of 25) behind title-level subscription access versus 15.4% (23 of 149) in the openly accessible remainder.Source: E. Structure of the evaluation gap · p. 15, Table XI · p. 15

Reader assessment by sequence

  • Among studies restricted to one sequence group, reader assessment was 21.7% for DWI/FLAIR without T1w/T2w and 11.9% for T1w/T2w without DWI/FLAIR, a descriptive difference that was not statistically significant.The comparison rests on 23 versus 84 studies (Fisher exact odds ratio 2.06, p = 0.31); the authors conclude there is no evidence readers are displaced from acute-pathology sequences.Source: E. Structure of the evaluation gap · p. 15

F. Methodological quality and risk of bias

Appraisal found the clearest risk in patient selection, while transparency omissions were widespread. The authors distinguish how these omissions affect performance estimates from how they constrain the review’s safety conclusion.

QUADAS-2 appraisal

  • 45/57 studies (78.9%) had low risk of bias for Flow and Timing, while Patient Selection was weakest, with 15/57 (26.3%) low risk and 23/57 (40.4%) high risk.Reference Standard and Index Test each had 26/57 (45.6%) low-risk ratings; applicability concerns were low for 30/57 (52.6%) at Index Test and 39/57 (68.4%) at Reference Standard.Source: F. Methodological quality and risk of… · pp. 15–16

Second-round appraisal

  • Second-round appraisal differed from the first round in all seven domains, with the largest shifts for Index Test risk and Reference Standard applicability.The pooled figures incorporate five studies re-appraised on recovered full text.Source: F. Methodological quality and risk of… · p. 16

Reporting transparency

  • 220/263 studies (83.7%) omitted at least one key methodological indicator; 109/263 lacked an identifiable reference standard, 54/263 omitted field strength, and 166/263 did not openly release code.The 109 without an identifiable reference standard included 47 reporting none and 62 naming an unspecified ground truth.Source: F. Methodological quality and risk of… · p. 16

Interpretation

  • The reporting and selection pattern is anti-conservative for reconstruction performance estimates but conservative for the review’s conclusion about the scarcity of safety-relevant evaluation.The authors state that the documented gap is therefore an understatement, not an estimate of actual deployment failures.Source: F. Methodological quality and risk of… · pp. 16–17

G. Requirements for safety-oriented evaluation

The review derives five requirements for safety-oriented evaluation, while concluding that no current evaluation paradigm settles reconstruction safety. The discussion also identifies a missing benchmark and access barriers that constrain meaningful evaluation.

G. Requirements for safety-oriented evaluation

  • Safety evaluation should pair global fidelity scores with focal task-based measures and same-data reader comparisons, use reference-independent validity, pre-specify acquisition-shift robustness, and detect hallucinations image-side.A benchmark containing the pathology at issue is a precondition for exercising several requirements.Source: G. Requirements for safety-oriented… · p. 17

V. DISCUSSION

  • Only 18 of 263 studies record fidelity metrics and reader assessment on the same data; reader evidence is mostly weak, condition-dependent inter-reader agreement.No included study reports a rank correlation between pixel-wise fidelity and radiologist diagnostic assessment.Source: §V · p. 17
  • Metric properties explain divergence: spatial averaging hides focal error, symmetric losses ignore diagnostic direction, adversarial terms reward invented texture, and motion can destroy registered references.Measurement-space residual checks are blind to null-space edits.Source: G. Requirements for safety-oriented… · p. 17, §V · p. 18
  • Exploratory analysis links the evaluation gap to a community split: only 4.9% of studies release code and run reader studies, while common benchmarks lack acute pathology.Engineering and clinical venues have different data access, making reader studies on available engineering datasets uninformative for lesion preservation.Source: §V · p. 18
  • Commercial DICOM exports omit inputs needed for prior-edit audits, so independent third-party auditing is structurally impossible without an open raw-data interface.Null-space projection needs the forward operator and coil sensitivities; counterfactual audits need controlled acquisition changes.Source: §V · p. 18
  • A lacunar infarct’s erasure changes global PSNR by about 0.03 dB under the stated error model, a figure papers could report without new data.The review reports that no paper currently provides this arithmetic.Source: §V · p. 18
  • No dataset combines raw k-space, acute lesions verified independently, lesion-level annotations, and redistribution rights, although each component exists separately.ISLES and ATLAS provide acute pathology, and fastMRI+ has reader-level pathology annotation, but included studies do not use them.Source: §V · p. 18

A. Priorities for future research

The review prioritizes paired metric–reader reporting, acquisition-shift testing, pathology-bearing benchmarks, and 3D anatomical coherence. Its conclusions are bounded by shallow, unaudited extraction, heterogeneous measures, and limited reader and prospective evidence.

A. Priorities for future research

  • Pair metric–reader reporting, make acquisition-shift robustness a primary endpoint, and use benchmarks containing the pathology at issue as priorities for future evaluation.The review also identifies pathology-bearing benchmarks as a precondition for several of its five requirements.Source: A. Priorities for future research · p. 19
  • Model 3D anatomical coherence while meeting scanner-side latency constraints, a mechanistic priority not measured in the review’s extraction.The authors note that slice-wise processing can introduce inter-slice discontinuities on isotropic 3D acquisitions.Source: A. Priorities for future research · p. 19

B. Threats to validity

  • Only 57 of 263 studies included any reader assessment, and 18 paired one with a metric, so the strongest conclusions rest on a small subset.The authors identify four threat classes that bound what follows from this evidence base.Source: B. Threats to validity · p. 19
  • Title, abstract, and vocabulary-based labels make prevalence counts floors, while the unverified metric label makes the 18/263 paired count downward-biased by an unknown amount.Only 84 studies were recorded as reporting a fidelity metric, and 119 lacked an evaluation-paradigm label.Source: B. Threats to validity · p. 19
  • Extraction has no reliability estimate; released quotations enable checking assignments but do not replace a statistical estimate of extraction reliability.The authors identify duplicate extraction on a random subsample as outstanding work.Source: B. Threats to validity · p. 19
  • Metric values grouped under one heading may represent different quantities of interest, limiting direct comparison of their correlations with reader scores.The venue classification is also a post-hoc keyword assignment.Source: B. Threats to validity · p. 19
  • Missing full texts and paywall differences constrain internal validity: reader assessment was more common in the inaccessible paywalled stratum than in openly accessible studies of that era.A sensitivity analysis on 258 fully verified studies left the findings directionally stable; appraisal agreement also leaves uncertainty in QUADAS-2 domain shares.Source: B. Threats to validity · pp. 19–20
  • Prospective multicentre reader evidence is thin, and no named corpus dataset is curated for acute stroke or intracranial haemorrhage.The review also cautions that retrospective, single-centre, preprocessed data can flatter performance relative to prospective multicentre use.Source: B. Threats to validity · p. 20

VI. CONCLUSION

The review finds weak agreement between conventional fidelity metrics and radiologist judgement, with paired evaluation rare and prevalence counts limited by extraction. It argues that pathology-bearing benchmarks and reporting conventions are needed to assess lesion preservation.

VI. CONCLUSION

  • Where metrics and readers were compared, conventional fidelity scores agreed weakly with radiologist judgement, and at most 18 of 263 studies recorded paired evaluation.The review states that no current evaluation paradigm alone answers the safety question.Source: §VI · p. 20
  • Only 4.9% of studies both released code and ran readers, and no named dataset was curated for acute stroke or intracranial haemorrhage.The proposed explanation involving community divisions is presented as a hypothesis, not an established structural fact.Source: §VI · p. 20
  • Pathology-bearing benchmarks and reporting conventions that never present a pixel-wise metric alone are needed to address whether reconstruction preserves small acute lesions.The review concludes that current measurements cannot support claims of diagnostic safety until both exist.Source: §VI · p. 21
  • A 100 mm3 infarct’s erasure shifts global PSNR by around 0.03 dB, below conventional reporting precision, under the stated error model.The authors propose reporting the smallest focal change the score can register at the stated acceleration and image size.Source: §VI · p. 21

FUNDING STATEMENT

The authors report institutional support, no external or industry funding, no need for ethics approval, and release review materials and analysis code. They also describe author contributions, manuscript-only AI assistance, and how to trace claims to study-level data.

FUNDING STATEMENT

  • The work received institutional research-computing and scholarship support from RMIT University, with no external, commercial, or industry funding.The funder had no role in the review or publication decision, and no MRI hardware or reconstruction-software vendor contributed support.Source: FUNDING STATEMENT · p. 22

ETHICS STATEMENT

  • The review used published, de-identified reports and collected no private health data, so institutional review board approval and informed consent were not required.Source: ETHICS STATEMENT · p. 22

DATA AND CODE AVAILABILITY

  • The protocol, screening decisions, extraction and QUADAS-2 worksheets, and analysis code are openly available on OSF and GitHub and archived at Zenodo.Included-study PDFs are not redistributed because of publisher licence terms.Source: DATA AND CODE AVAILABILITY · p. 22

CREDIT AUTHORSHIP CONTRIBUTION STATEMENT

  • Dat Tat Mai led conceptualization, methodology, software, investigation, data curation, analysis, visualization, administration, and writing; the other authors contributed investigation, validation, supervision, resources, and writing.Source: CREDIT AUTHORSHIP CONTRIBUTION STATEMENT · p. 22

TECHNOLOGIES IN THE MANUSCRIPT PREPARATION

TRACE A CLAIM TO THEM

  • Study-level extracted characteristics and supporting quotations in included_characteristics.csv let readers reproduce counts and identify contributing studies.Supplementary Tables S1–S3 index the 263 studies, and the supplementary characteristics file adds QUADAS-2 and reproducibility ratings.Source: TRACE A CLAIM TO THEM · p. 22

APPENDIX B. SUPPLEMENTARY MATERIAL

The supplementary document provides study lists, comparison tables, search and extraction materials, and supporting details for the review’s thematic and quantitative analyses.

APPENDIX B. SUPPLEMENTARY MATERIAL

  • Supplementary Tables S1–S12 and Figure S1 include the 263-study bibliography, related-review comparisons, search and extraction materials, and evidence supporting the review’s analyses.Tables S10–S12 provide supporting detail on fastMRI+ annotation, sequence-strata correction, and venue cross-tabulation by era.Source: Appendix B · pp. 22–23

Related papers

Loading 2608.28714v1…