A versatile foundation model for cine cardiac magnetic resonance image analysis tasks

Anish N Bhuva; Charlotte Manisty; James C Moon; Matthew J Clarkson; Rhodri Huw Davies; Thomas A Treibel; Weixi Yi; Wenjia Bai; Yipeng Hu; Yunguan Fu

REVIEW 3 major objections 8 minor 2 cited by

A versatile foundation model for cine cardiac magnetic resonance image analysis tasks

T0 review · 3 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read CineMA, a masked-autoencoder foundation model pre-trained on 15 million cine cardiac MR images, claims to consistently outperform conventional CNNs in ventricular segmentation and ejection fraction estimation, and to match or beat them…

desk verdict CineMA is a large, publicly released cine CMR foundation model with mostly convincing gains over simple CNNs—the key caveats are weak baselines and an abstract that overstates mortality prediction. read the letter →

arxiv 2506.00679 v2 pith:RXP6LBRR submitted 2025-05-31 eess.IV cs.AIcs.CV

Yunguan Fu , Wenjia Bai , Weixi Yi , Charlotte Manisty , Anish N Bhuva , Thomas A Treibel , James C Moon , Matthew J Clarkson

show 2 more authors

Rhodri Huw Davies Yipeng Hu

This is my paper · ORCID

classification eess.IVcs.AIcs.CV

keywords foundationmodelcardiacmagneticresonanceimagingmaskedautoencoderself-supervisedlearningcineCMRventricularsegmentationejectionfractiondiseasedetection

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 argues that a single foundation model pre-trained on large-scale cine cardiac MRI can replace task-specific convolutional networks across the main clinical image-analysis tasks: delineating the ventricles, estimating ejection fraction, detecting disease, and locating anatomical landmarks. The model, CineMA, is a masked autoencoder with convolutional and transformer components trained on 15 million images from 74,916 UK Biobank studies, then fine-tuned per task. Across eight independent datasets, it reports consistently higher Dice scores and lower ejection-fraction errors than randomly initialised UNet and ResNet-50 models, with the advantage persisting when only half of the fine-tuning labels are used. The paper also reports that CineMA detects cardiac functional changes associated with diabetes, hypertension, and cancer, predicts mortality, and shows small demographic disparities. If these results hold, pre-trained cardiac imaging models would reduce annotation burden and improve generalisation for clinical and epidemiological use.

What carries the argument

The central object is a multi-view convolution-transformer masked autoencoder: each of the four cine views (short-axis and two-, three-, and four-chamber long-axis) is independently encoded by convolutional layers, 75% of patches are randomly masked, visible tokens from all views pass through a shared Vision Transformer encoder, and a shared transformer decoder reconstructs the masked patches under a mean squared error loss. This self-supervised reconstruction task is what transfers anatomy-aware representations to downstream tasks. For fine-tuning, the decoder is discarded, only the encoder branches of the available views are kept, and task-specific heads are added: UNetR-style decoders for segmentation and landmark heatmaps, and linear prediction layers for tabular labels such as ejection fraction, disease class, sex, and vendor.

What would settle it

Re-run the benchmark on ACDC, M&Ms, and M&Ms2 with a per-dataset auto-configured segmentation method such as nnU-Net, which the paper excludes in Section 4.11; if that method matches or beats CineMA on Dice score and ejection-fraction error, the claimed consistent superiority over CNNs would no longer hold.

Watch

Extended reading notes

Core claim

The central claim is that CineMA, a multi-view convolution-transformer masked autoencoder pre-trained on over 15 million cine CMR images from 74,916 UK Biobank subjects, is a versatile foundation model whose fine-tuned variants consistently outperform conventional CNNs trained from scratch in delineating the left and right ventricles and myocardium, and in estimating ejection fraction, a key measure of cardiac function. The paper further claims that CineMA surpasses CNNs in disease detection, achieves higher specificity while maintaining sensitivity, matches CNNs on long-axis functional measures such as MAPSE and GLS, and preserves its advantage with half the training data and under population shift on external Kaggle and Rescan cohorts. The authors hold that this is the first cine CMR foundation model to demonstrate superior performance to CNNs and previous methods across such a broad range of tasks, and that CineMA-derived metrics carry clinical meaning through associations with diabetes, hypertension, cancer, and survival, with consistent performance across demographic subgroups.

Load-bearing premise

The load-bearing premise is that the fixed-hyperparameter UNet and ResNet-50 baselines are strong enough representatives of conventional CNNs; if they were undertrained or under-tuned, the reported advantage could reflect weak comparisons rather than the value of pre-training.

Editorial extensions

If this is right

Cardiac image analysis could shift from per-task models trained from scratch to a single pre-trained backbone fine-tuned per task, reducing the clinical annotation burden.
With half the fine-tuning labels, CineMA matches the ejection-fraction accuracy of a fully trained UNet, so institutions with limited annotated data could still deploy accurate function assessment.
EF estimation on external Kaggle and Rescan cohorts improves despite a marked population shift, supporting zero-shot deployment on new scanners and disease mixes.
Higher specificity in disease detection with maintained sensitivity suggests automated screening could reduce false-positive referrals while keeping false negatives low.
CineMA-predicted left ventricular ejection fraction is associated with one-, two-, and five-year mortality at effect sizes comparable to UK Biobank-provided LVEF, giving automated metrics prognostic value.

Reading between the lines

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

Because pre-training used only UK Biobank cine images, the transfer gains are likely bounded by that population and acquisition protocol; adding external cine data and modalities such as LGE or T1 mapping is a natural test of whether the current LGE/T1 results, where CineMA only matched CNNs, would improve.
One could ablate the multi-view design by pre-training separate single-view masked autoencoders at the same scale; if multi-view reconstruction is the active ingredient, it should show a measurable advantage over scale alone.
The fairness analysis assumes equal positive-outcome rates across groups; where ground-truth labels exist, per-group error analysis could check whether disparity ratios close to 1 reflect true accuracy rather than correlated bias.

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

CineMA is a solid, workmanlike foundation-model paper, and the central empirical claim—that large-scale masked-autoencoder pre-training helps downstream cine CMR tasks—mostly survives contact with the data. What it is not is a demonstration that a foundation model beats strong CNN baselines across the board, and the abstract's "can also predict mortality" overstates what the survival analysis actually does.

The genuinely useful parts: a 74,916-subject, 15-million-image pre-training corpus, a multi-view architecture that handles SAX and LAX in one backbone, and a benchmark spanning eight datasets with public code and weights. The random-init same-architecture control is a good design choice; it separates pre-training benefit from architecture capacity. The zero-shot EF numbers on Kaggle and Rescan, with repeated-scan coefficient of variation, are a real plus and consistent with the pre-training story. The disease-association and Cox survival analyses are honest secondary analyses, and the Limitations section says plainly that they did not model or predict clinical outcomes.

The soft spots are in the comparison. The CNN baselines are a five-layer UNet and ResNet-50 trained with the same fixed hyperparameter set, and nnUNet is excluded by construction. That makes "consistently outperformed conventional CNNs" carry more weight than it can bear: on MAPSE/GLS, EMIDEC, and MyoPS2020, CineMA matches UNet, and CineMARandInit is often worse than UNet. The segmentation gains are real but modest—a few Dice points, mostly on the right ventricle. For the strongest claim, the authors need either a properly tuned CNN baseline or softened language.

The survival framing also needs fixing. The paper fits a Cox model with CineMA-predicted LVEF as an input; that is an association study, not mortality prediction. The abstract and intro should say so.

The statistics are loosely described (bootstrapping then two-sided t-tests), and the demographic fairness metric is simple, but neither is fatal. Using the same UKB cohort for pre-training and the association analyses is a potential circularity concern for those secondary analyses, though the main task benchmarks are external and the labels were not used in pre-training, so it is not load-bearing. The related-work citations look fair—they name Shad, Jacob, Zhang and position against them.

Bottom line: I would send this to peer review. It is a useful resource and a fairly run benchmark, but it needs a major revision to either strengthen the baselines or narrow the claims before the headline can be accepted.

Referee Report

3 major / 8 minor

Summary. This paper presents CineMA, a multi-view convolution-transformer masked autoencoder pre-trained on 74,916 UK Biobank cine CMR studies (over 15M images), and benchmarks fine-tuned variants against from-scratch CNNs (UNet and ResNet-50) and a same-architecture random-initialisation control across segmentation, ejection-fraction estimation, landmark localisation, disease classification, long-axis function measurement, and additional association, survival, and fairness analyses. The central claim is that fine-tuned CineMA consistently outperforms conventional CNNs across a broad range of cine CMR tasks, with particular benefits in low-data and population-shift settings, and that it is the first foundation model to achieve such broad superiority.

Significance. If the central claim holds, this is a substantial contribution to medical imaging foundation models. The strengths include the large pre-training corpus, the unified multi-view architecture, external validation across eight datasets, the use of a random-initialisation same-architecture control, and the public release of code and model weights. The additional analyses linking CineMA-derived metrics to systemic disease, survival, and demographic subgroups are valuable and go beyond typical benchmark studies. However, the strength of the 'consistently outperformed CNNs' claim currently depends on the adequacy of the CNN baselines, on the validity of the significance-testing procedure, and on internal consistency of the reported numbers; these issues need to be resolved before the headline claim is fully supported.

major comments (3)

[Sections 4.10 and 4.11, Table 16] The central claim of consistent superiority over conventional CNNs rests on the fixed hyperparameter set in Table 16 and on the exclusion of nnUNet. The paper states that all models share one set of hyperparameters 'not tuned via exhaustive search' and that nnUNet is omitted because fair comparison with an AutoML pipeline remains an open research direction. This is a load-bearing issue because CineMARandInit, the same-architecture control, underperforms UNet on several datasets (e.g., M&Ms2 SAX mean Dice 85.97 vs 88.02 in Table 2), so the raw comparison may conflate pre-training benefit with hyperparameter suitability for the transformer. I recommend either adding a properly tuned CNN baseline (e.g., nnUNet for segmentation and a tuned ResNet for tabular tasks) or providing evidence that the fixed recipe is near-optimal for the CNN baselines, such as learning curves and a small hyperparameter sensitivity scan.
[Section 4.12, Tables 2-9, Figures 2-3] The statistical significance claims are based on two-sided t-tests applied to values obtained by bootstrapping the test set. Bootstrap samples are not independent, so the t-test's independence assumption is violated, and the resulting p-values can be misleading. The procedure also averages over the three random seeds within each bootstrap sample, so seed variability is not propagated into the significance test. Please replace this with a valid procedure, such as bootstrap percentile or BCa confidence intervals for the metric differences, or a paired permutation test that resamples over both test examples and random seeds. This is important because several claims of 'significantly higher' performance depend on these p-values.
[Tables 3 and 4, Section 2.2] The claim of consistent superiority in EF estimation is not fully supported by the reported numbers. In Table 3, the M&Ms2 (SAX) LVEF MAE is 3.53% for CineMA FineTune versus 3.50% for UNetRandInit. In Table 4, the model trained on M&Ms gives a Rescan LVEF MAE of 4.08% for CineMA versus 3.90% for UNet, and a coefficient of variation of 7.84% versus 7.65%. These are cases where CineMA does not outperform the CNN baseline. The abstract and Section 2.4 use 'consistently outperformed' as a headline, which is too strong in light of these entries. Please either qualify the claim as task- and dataset-dependent, or provide statistical evidence that these small adverse differences are within noise and that the overall pattern remains favourable.

minor comments (8)

[Table 1 and Section 4.5] The Landmark dataset is listed as n=2,888 in Table 1 but Section 4.5 describes 1,444 studies; please clarify whether the table counts studies, images, or image-view instances.
[Table 2 caption] The caption contains a typo: 'myocaridum' should be 'myocardium'.
[Table 6 caption] The caption states that CineMA FineTune 'outperformed or matched UNet', but for M&Ms2 LAX 4C vendor classification the ROC AUC is 77.76% for CineMA versus 84.41% for ResNetRandInit, which is worse; please correct the caption or the claim.
[Section 4.7] The description of the survival label derivation is unclear: if age at death is not recorded, participants are assumed to have survived, but the use of 'last hospital visit occurred in 2022 or earlier' to define 1-year survival needs a more explicit censoring rule, especially for participants whose last visit was before 2022.
[Figure 2 caption] The caption repeats 'R V' and 'R V,' and the abbreviations are not expanded consistently; please revise for clarity.
[Section 4.12] Bootstrapping is repeated only 100 times; for stable confidence intervals and p-values, 1000 or more bootstrap repetitions are typical and would be preferable.
[Section 2.5 and Table 5] The pooled specificity and sensitivity values 60.06%/31.55% quoted in the text are not directly derivable from the per-dataset rows in Table 5; please state explicitly how the pooling was performed.
[Section 4.9] The sentence 'A fixed proportion (75%) of patches is randomly masked' should specify whether this proportion applies per view and whether the mask ratio is the same for LAX and SAX views.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: CineMA's central benchmark claims rest on external task labels and held-out datasets, with pre-training benefit isolated by a random-init architecture control.

full rationale

The paper's central claims are empirical benchmark results rather than derivations, and I found no step where a predicted quantity is defined in terms of a fitted input or where a load-bearing premise reduces to a self-citation. Pre-training is self-supervised on unlabeled UKB cine frames; downstream segmentation, EF, disease, landmark, and survival evaluations use independent labels (ACDC, M&Ms, M&Ms2, Kaggle, Rescan, Landmark, EMIDEC, MyoPS2020) or UKB disease/survival fields that were not used in pre-training. The Kaggle/Rescan zero-shot EF result is computed from segmentations and compared against held-out labels, not training labels. The association and survival analyses are epidemiological uses of predicted metrics, not predictions validated on their own fitting targets. The random-init architecture control (CineMARandInit) explicitly separates pre-training benefit from architecture capacity. The only notable concern is experimental fairness: Section 4.11 omits nnUNet and Section 4.10 fixes all hyperparameters 'not tuned via exhaustive search,' so the CNN baselines may be weak; however, this is a benchmarking adequacy concern, not circularity, because no result is forced by construction. The citation to Bhuva et al. [6] for Rescan data and prior CV is a minor self-citation but not load-bearing: it supplies an external comparison value, not the derivation of CineMA's performance.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced. The central claim rests on the transferability of self-supervised representations, which depends on several dataset and modeling assumptions listed above. The most fragile assumption is that the chosen CNN baselines and fixed hyperparameters provide a fair comparison, since nnUNet was explicitly excluded.

free parameters (4)

mask ratio = 0.75
Chosen following MAE (He et al.), fixed for all pre-training. A hyperparameter not fitted to downstream data.
model size (base config) = 126M parameters
Adopted from He et al. base ViT-Large config. Not tuned.
Gaussian sigma for landmark heatmaps = 3
Chosen for heatmap generation in Section 4.5; affects landmark localisation loss but not central pre-training claim.
downstream hyperparameters = see Table 16
Fixed across experiments and stated as not exhaustively tuned; they could influence the comparison with baselines.

assumptions (6)

domain assumption UK Biobank cine CMR images are representative of clinical cine CMR and pre-training on this population transfers to other populations (Kaggle, Rescan).
Used in pre-training and zero-shot evaluation; if the population shift is too large, the gains may not generalize.
domain assumption The Landmark dataset labels (mitral valve and apex) are accurate enough to serve as ground truth for long-axis function.
Section 4.5; if labels are noisy, the MAPSE/GLS comparisons are affected.
domain assumption The max/min left ventricle volume over the cardiac cycle correctly identifies end-diastole and end-systole for EF calculation in zero-shot evaluation.
Section 2.3; assumes no arrhythmia or tracking errors.
ad hoc to paper A fair model should predict equal rates of LVEF > threshold across ethnic groups.
Section 2.8 states this assumption; it ignores possible true differences in cardiac function between groups.
domain assumption The Cox proportional hazards model adjusted for age, sex, and BMI adequately controls confounding in the survival association.
Section 4.13; unmeasured confounders could bias the association.
ad hoc to paper The fixed hyperparameters in Table 16 are suitable for both CineMA and the CNN baselines.
Section 4.10/4.11; if the same hyperparameters favor transformer fine-tuning over CNN random init, the comparison is biased.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A versatile foundation model for cine cardiac magnetic resonance image analysis tasks." pith.science (2026). https://pith.science/paper/RXP6LBRR

@misc{pith2026250600679,
  author       = {Pith},
  title        = {Pith review of: A versatile foundation model for cine cardiac magnetic resonance image analysis tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RXP6LBRR}},
  note         = {Machine review of arXiv:2506.00679}
}

read the original abstract

Here we present a versatile foundation model that can perform a range of clinically-relevant image analysis tasks, including segmentation, landmark localisation, diagnosis, and prognostication. A multi-view convolution-transformer masked autoencoder, named as CineMA, was trained on 15 million cine images from 74,916 subjects. The model was validated on multiple image analysis tasks and compared to existing models on >4,500 images from eight independent datasets with diverse population characteristics, representing the largest benchmark study for cine CMR so far. CineMA consistently outperformed conventional convolutional neural networks (CNNs) in delineating ventricular boundaries and estimating ejection fraction, a key measure of cardiac function. The improved performance was preserved, even when the model only used half of fine-tuning data. CineMA also surpassed CNNs in disease detection and matched their performance in long-axis function measurement. Interestingly, we found that CineMA can also detect cardiac changes in systemic diseases, such as diabetes, hypertension and cancer, and can also predict mortality. Finally, we assessed model fairness and demonstrated consistent model performance across demographic subgroups. These findings highlight CineMA's accuracy, learning efficiency, adaptability, and fairness, underscoring its potential as a foundation model for automated cardiac image analysis to support clinical workflow and cardiovascular research. All training and inference code and models are made publicly available at https://github.com/mathpluscode/CineMA.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Development of an automated, reliable, and clinically meaningful artificial intelligence (AI) tool for diagnosing cardiac disease from conventional cardiovascular magnetic resonance (CMR) images
cs.CV 2026-07 conditional novelty 6.0 of 10

A five-class CMR disease classifier built from LLM-derived report labels and three fine-tuned vision foundation models reached ensemble AUCs of 0.84–0.97 on a single-center test set.
Analysis of Image-and-Text Uncertainty Propagation in Multimodal Large Language Models with Cardiac MR-Based Applications
cs.CV 2025-07 conditional novelty 5.0 of 10

A linear uncertainty-propagation model fitted on cardiac MRI plus health-record text is shown to transfer across prediction tasks and data distributions, enabling cheaper uncertainty estimates.

Reference graph

Works this paper leans on

32 extracted references · 23 canonical work pages · cited by 2 Pith papers

[1]

A., Fuster, V

Mensah, G. A., Fuster, V. & Roth, G. A. A heart-healthy and stroke-free world: using data to inform global action (2023)

work page 2023
[2]

S., Fran¸ cois, C

Rajiah, P. S., Fran¸ cois, C. J. & Leiner, T. Cardiac mri: state of the art.Radiology 307, e223008 (2023)

work page 2023
[3]

& Schulz-Menger, J

von Knobelsdorff-Brenkenhoff, F. & Schulz-Menger, J. Cardiovascular magnetic resonance in the guidelines of the european society of cardiology: a comprehensive summary and update. Journal of Cardiovascular Magnetic Resonance 25, 42 (2023)

work page 2023
[4]

Schulz-Menger, J. et al. 2025 esc guidelines for the management of myocarditis and pericarditis: Developed by the task force for the management of myocarditis and pericarditis of the european society of cardiology (esc)endorsed by the associ- ation for european paediatric and congenital cardiology (aepc) and the european association for cardio-thoracic su...

work page doi:10.1093/eurheartj/ehaf192 2025
[5]

Kramer, C. M. et al. Standardized cardiovascular magnetic resonance imaging (cmr) protocols: 2020 update. Journal of Cardiovascular Magnetic Resonance 22, 17 (2020)

work page 2020
[6]

Bhuva, A. N. et al. A multicenter, scan-rescan, human and machine learning cmr study to test generalizability and precision in imaging biomarker analysis. Circulation: Cardiovascular Imaging 12, e009214 (2019)

work page 2019
[7]

Bernard, O. et al. Deep learning techniques for automatic mri cardiac multi- structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging 37, 2514–2525 (2018)

work page 2018
[8]

Campello, V. M. et al. Multi-centre, multi-vendor and multi-disease cardiac seg- mentation: the m&ms challenge. IEEE Transactions on Medical Imaging 40, 3543–3554 (2021)

work page 2021

Show all 32 references

[9]

Deep learning segmentation of the right ventricle in cardiac mri: the m&ms challenge

Mart ´ ın-Isla, C.et al. Deep learning segmentation of the right ventricle in cardiac mri: the m&ms challenge. IEEE Journal of Biomedical and Health Informatics 27, 3302–3313 (2023)

2023
[10]

A, N. et al. Second annual data science bowl (2015). URL https://kaggle.com/ competitions/second-annual-data-science-bowl

2015
[11]

Xue, H. et al. Landmark detection in cardiac mri by using a convolutional neural network. Radiology: Artificial Intelligence 3, e200197 (2021). 31

2021
[12]

Lalande, A. et al. Emidec: a database usable for the automatic evaluation of myocardial infarction from delayed-enhancement cardiac mri. Data 5, 89 (2020)

2020
[13]

Li, L. et al. Myops: A benchmark of myocardial pathology segmentation combin- ing three-sequence cardiac magnetic resonance images. Medical Image Analysis 87, 102808 (2023)

2023
[14]

Bommasani, R. et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)

2021 arXiv
[15]

Ma, J. et al. Segment anything in medical images. Nature Communications 15, 654 (2024)

2024
[16]

& Wang, L

Sun, Y., Wang, L., Li, G., Lin, W. & Wang, L. A foundation model for enhanc- ing magnetic resonance images and downstream segmentation, registration and diagnostic tasks. Nature Biomedical Engineering 9, 521–538 (2025)

2025
[17]

Shad, R. et al. A generalizable deep learning system for cardiac mri. arXiv preprint arXiv:2312.00357 (2023)

2023
[18]

Jacob, A. J. et al. Towards a vision foundation model for comprehensive assessment of cardiac mri. arXiv preprint arXiv:2410.01665 (2024)

2024
[19]

Zhang, Y. et al. Towards cardiac mri foundation models: Comprehensive visual- tabular representations for whole-heart assessment and beyond. arXiv preprint arXiv:2504.13037 (2025)

2025 arXiv
[20]

Caron, M. et al. Emerging properties in self-supervised vision transformers (2021)

2021
[21]

He, K. et al. Masked autoencoders are scalable vision learners. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022)

2022
[22]

Zhou, Y. et al. A foundation model for generalizable disease detection from retinal images. Nature 622, 156–163 (2023)

2023
[23]

& Zamir, A

Bachmann, R., Mizrahi, D., Atanov, A. & Zamir, A. Multimae: Multi-modal multi-task masked autoencoders (2022)

2022
[24]

Gao, P. et al. Convmae: Masked convolution meets masked autoencoders. arXiv preprint arXiv:2205.03892 (2022)

2022 arXiv
[25]

C., Neubauer, S

Raisi-Estabragh, Z., Harvey, N. C., Neubauer, S. & Petersen, S. E. Cardiovascu- lar magnetic resonance imaging in the uk biobank: a major international health research resource. European Heart Journal-Cardiovascular Imaging 22, 251–258 (2021)

2021
[26]

Hatamizadeh, A. et al. Unetr: Transformers for 3d medical image segmentation (2022). 32

2022
[27]

& Lyman, K

Targ, S., Almeida, D. & Lyman, K. Resnet in resnet: Generalizing residual architectures. arXiv preprint arXiv:1603.08029 (2016)

2016 arXiv
[28]

Seyyed-Kalantari, L., Liu, G., McDermott, M., Chen, I. Y. & Ghassemi, M. Chex- clusion: Fairness gaps in deep chest x-ray classifiers. BIOCOMPUTING 2021: proceedings of the Pacific symposium 232–243 (2020)

2020
[29]

Dai, D. et al. Deepseekmoe: Towards ultimate expert specialization in mixture- of-experts language models. arXiv preprint arXiv:2401.06066 (2024)

2024 arXiv
[30]

F., Kohl, S

Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J. & Maier-Hein, K. H. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods 18, 203–211 (2021)

2021
[31]

McIntosh-Smith, S., Alam, S. R. & Woods, C. Isambard-ai: a leadership class supercomputer optimised specifically for artificial intelligence. arXiv preprint arXiv:2410.11199 (2024)

2024 arXiv
[32]

& Wu, X.-J

Feng, Z.-H., Kittler, J., Awais, M., Huber, P. & Wu, X.-J. Wing loss for robust facial landmark localisation with convolutional neural networks (2018). 33

2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

A., Fuster, V

Mensah, G. A., Fuster, V. & Roth, G. A. A heart-healthy and stroke-free world: using data to inform global action (2023)

work page 2023

[2] [2]

S., Fran¸ cois, C

Rajiah, P. S., Fran¸ cois, C. J. & Leiner, T. Cardiac mri: state of the art.Radiology 307, e223008 (2023)

work page 2023

[3] [3]

& Schulz-Menger, J

von Knobelsdorff-Brenkenhoff, F. & Schulz-Menger, J. Cardiovascular magnetic resonance in the guidelines of the european society of cardiology: a comprehensive summary and update. Journal of Cardiovascular Magnetic Resonance 25, 42 (2023)

work page 2023

[4] [4]

Schulz-Menger, J. et al. 2025 esc guidelines for the management of myocarditis and pericarditis: Developed by the task force for the management of myocarditis and pericarditis of the european society of cardiology (esc)endorsed by the associ- ation for european paediatric and congenital cardiology (aepc) and the european association for cardio-thoracic su...

work page doi:10.1093/eurheartj/ehaf192 2025

[5] [5]

Kramer, C. M. et al. Standardized cardiovascular magnetic resonance imaging (cmr) protocols: 2020 update. Journal of Cardiovascular Magnetic Resonance 22, 17 (2020)

work page 2020

[6] [6]

Bhuva, A. N. et al. A multicenter, scan-rescan, human and machine learning cmr study to test generalizability and precision in imaging biomarker analysis. Circulation: Cardiovascular Imaging 12, e009214 (2019)

work page 2019

[7] [7]

Bernard, O. et al. Deep learning techniques for automatic mri cardiac multi- structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging 37, 2514–2525 (2018)

work page 2018

[8] [8]

Campello, V. M. et al. Multi-centre, multi-vendor and multi-disease cardiac seg- mentation: the m&ms challenge. IEEE Transactions on Medical Imaging 40, 3543–3554 (2021)

work page 2021