A Latent ODE Approach to Spatiotemporal Modeling of Cine Cardiac MRI

David Br\"uggemann; Ekaterina Krymova; Firat \"Ozdemir; Jochen von Spiczak; Mathieu Salzmann; Olga V. Demler; Robert Manka; Samia Mora; Sebastian Kozerke

arxiv: 2606.26718 · v1 · pith:4S2ZZ6G3new · submitted 2026-06-25 · 💻 cs.AI · cs.CV

A Latent ODE Approach to Spatiotemporal Modeling of Cine Cardiac MRI

David Br\"uggemann , Ekaterina Krymova , Firat \"Ozdemir , Jochen von Spiczak , Sebastian Kozerke , Samia Mora , Robert Manka , Mathieu Salzmann

show 1 more author

Olga V. Demler

This is my paper

Pith reviewed 2026-06-26 04:45 UTC · model grok-4.3

classification 💻 cs.AI cs.CV

keywords latent dynamical modelneural ODEcardiac MRIheart failure predictionspatiotemporal modelinggraph autoencoderUK Biobankprognostic score

0 comments

The pith

A latent dynamical model encodes full-cycle cardiac MRI motion as a continuous trajectory whose deviations predict incident heart failure better than standard markers.

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

The paper builds a model that turns cine CMR images of the ventricles into a smooth path through a hidden space, using differential equations tuned to heart rate and a graph mesh decoder to keep the motion anatomically consistent. It then measures how far each person's trajectory strays from an expected starting state set by basic covariates, and feeds that distance into a survival model. In 72,386 UK Biobank participants the resulting score raised the C-index of refitted risk equations from 0.704 to 0.785, beating the 0.764 obtained from seven conventional cardiac measurements. The same model also produced more faithful image reconstructions and more realistic motion sequences than versions that dropped the graph structure or the ODE component. These results indicate that the entire motion cycle carries risk information that snapshot summaries miss.

Core claim

Encoding bi-ventricular anatomy and full-cycle cine motion as a continuous latent trajectory, using heart-rate-aware neural ODE dynamics and a graph-based mesh autoencoder, produces deviations from a covariate-conditioned prior that independently predict incident heart failure in a large population cohort.

What carries the argument

Heart-rate-aware neural ordinary differential equation dynamics paired with a graph-based mesh autoencoder that represents 3D+t ventricular motion as a continuous latent trajectory.

If this is right

Adding the latent score to refitted pooled cohort equations raises the stratified C-index from 0.704 to 0.785 in held-out UK Biobank data.
The full model achieves a better balance of reconstruction fidelity, generative realism, and prognostic performance than non-graph or non-ODE variants.
Continuous full-cycle ventricular motion supplies cardiac phenotypes beyond those obtained from selected cardiac phases.
The latent score supplies incremental information relative to seven established cardiac markers.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same latent trajectory could be tested as a predictor for other cardiovascular events such as arrhythmia or stroke.
Combining the latent score with genetic or lifestyle data might yield further gains in risk stratification.
Prospective clinical studies would be needed to determine whether the score changes patient management decisions.

Load-bearing premise

The learned latent deviations capture prognostic information that is independent of seven established cardiac markers and that the UK Biobank-trained model applies to other populations.

What would settle it

An external validation cohort in which adding the latent score to the refitted pooled cohort equations produces no improvement in stratified C-index beyond the value achieved by the seven conventional markers alone.

Figures

Figures reproduced from arXiv: 2606.26718 by David Br\"uggemann, Ekaterina Krymova, Firat \"Ozdemir, Jochen von Spiczak, Mathieu Salzmann, Olga V. Demler, Robert Manka, Samia Mora, Sebastian Kozerke.

**Figure 1.** Figure 1: Framework overview. A GNN–Transformer encodes bi-ventricular mesh sequences and subject covariates into a latent vector 𝐳. A neural ODE evolves this state continuously, and a graph decoder reconstructs anatomically consistent cardiac motion 𝐱̂(𝑡). Survival analysis links 𝐳 to future outcomes to produce a personalized heart failure risk score. interpretable but reduce the data twice: they compress highdime… view at source ↗

**Figure 2.** Figure 2: Probabilistic framework for bi-ventricular 3D+t cardiac motion. Left: Frame-wise graph neural networks (GNNs) encode a mesh sequence 𝐱(𝑡1 ), …, 𝐱(𝑡𝑇 ); heart-rate-aware phase embeddings and a temporal transformer infer an approximate posterior over the end-diastolic latent state 𝐳(0). Middle: An MLP predicts a covariate-conditioned prior on 𝐳(0), and KL regularization aligns posterior and prior while prese… view at source ↗

**Figure 3.** Figure 3: Sex-stratified C-indices in evalution cohort for secondary cardiovascular endpoints, using the conventional CMRindex score 𝑟conv or the latent score 𝑟lat, in addition to PCP-HF variables [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 5.** Figure 5: Comparison of the covariate-conditioned prior-mean and posterior-mean reconstructions for two representative high-risk female subjects. Left: preserved ejection fraction. Right: reduced ejection fraction. Top row: LV volume curves. Bottom row: LV wall thickness. Blue denotes the prior mean and orange the posterior mean inferred from the observed mesh sequence. In the preserved-EF case, the posterior is dom… view at source ↗

**Figure 6.** Figure 6: Surface displacement of the mesh decoded from the posterior mean relative to the subject-specific prior mean for the two female subjects in [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

**Figure 8.** Figure 8: To evaluate latent robustness we compare acrossseed representational similarity for the full model (left) and the ablation using a standard normal prior (right). 16 32 64 128 256 ( -dimension) 1.4 1.6 1.8 Rec. error (mm) 1.25 2.5 5 10 20  (KL weight) 0.775 0.780 0.785 C-index Rec. error C-index [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗

**Figure 7.** Figure 7: Component analysis of the proposed model. Panel a ablates important model components for HF prediction. Panel b visualizes the learned phase trajectories and shows improved alignment of physiological phase across subjects after warping. ED: end diastole, ES: end systole, HR: heart rate, LV: left ventricle. the latent score reached a C-index of 0.785 in a held-out UK Biobank evaluation cohort. Across archit… view at source ↗

read the original abstract

Cardiac magnetic resonance imaging (CMR) captures rich spatiotemporal information about ventricular structure and motion, but conventional risk models use only a few image-derived indices from selected cardiac phases. We present a latent dynamical model that encodes bi-ventricular anatomy and full-cycle cine motion as a continuous latent trajectory, using heart-rate-aware neural ordinary differential equation (ODE) dynamics and a graph-based mesh autoencoder to reconstruct anatomically consistent 3D+t ventricular motion. A covariate-conditioned prior defines the expected end-diastolic latent state, and a Cox proportional hazards model tests whether deviations from this prior predict incident heart failure. We studied 72,386 UK Biobank participants without baseline cardiovascular disease, including 367 incident heart failure events. In a held-out evaluation subset, adding the latent score to refitted pooled cohort equations improved the stratified C-index from 0.704 to 0.785, compared with 0.764 for seven established cardiac markers. Compared with non-graph and non-ODE approaches, the proposed model gave the best trade-off between reconstruction fidelity, generative realism, and downstream prognostic performance. These results suggest that continuous full-cycle modeling of ventricular motion provides informative cardiac phenotypes beyond conventional CMR summaries, while external validation in more representative patient cohorts is required before clinical risk-prediction use.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The latent ODE plus graph autoencoder pipeline produces a C-index lift on UK Biobank heart failure data, but the gain's independence from the seven standard markers is not demonstrated in the abstract.

read the letter

The paper introduces a heart-rate-aware neural ODE combined with a graph mesh autoencoder to encode full-cycle bi-ventricular cine MRI as a continuous latent trajectory, then feeds deviations from a covariate-conditioned prior into a Cox model. This specific combination for 3D+t reconstruction and downstream risk prediction does not appear in the referenced prior work.

On the 72k UK Biobank participants it reports better reconstruction fidelity and generative realism than non-graph or non-ODE baselines, and the latent score raises stratified C-index from 0.704 to 0.785 when added to refitted pooled cohort equations, versus 0.764 from seven established cardiac markers.

The main soft spot is the missing evidence that the latent deviations carry information independent of those seven markers already extractable from the same images. No correlation table or ablation removing the markers is mentioned, and the abstract gives no explicit statement that the autoencoder and ODE were fit only on the training partition before the held-out Cox evaluation. The circularity risk is therefore real until the methods are checked.

External validation is correctly flagged as needed; the UK Biobank-only result is not yet transportable.

This is for cardiac imaging and risk-modeling groups. The technical approach and sample size are solid enough to merit referee time even if the incremental clinical value turns out modest.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces a latent dynamical model that encodes bi-ventricular anatomy and full-cycle cine cardiac MRI motion as a continuous latent trajectory via a heart-rate-aware neural ODE and graph-based mesh autoencoder. A covariate-conditioned prior is used to define expected end-diastolic states, with deviations from this prior tested as predictors of incident heart failure via Cox regression in 72,386 UK Biobank participants (367 events). In a held-out subset, adding the latent score to refitted pooled cohort equations raises the stratified C-index from 0.704 to 0.785, outperforming seven established cardiac markers (0.764); the model also shows the best trade-off versus non-graph and non-ODE baselines in reconstruction and generative metrics.

Significance. If the reported incremental value holds after confirming independence and no leakage, the work would demonstrate that continuous full-cycle latent trajectories can extract prognostic cardiac phenotypes beyond static CMR summaries, with potential to improve risk models. Credit is due for the explicit comparison to non-graph/non-ODE ablations, the large cohort, and the abstract's clear statement that external validation is required before clinical use. The technical integration of graph autoencoders with ODE dynamics for anatomically consistent 3D+t reconstruction is a clear strength.

major comments (2)

[Abstract] Abstract: the headline claim that the latent score adds incremental value (C-index rising from 0.704 to 0.785 versus 0.764 for the seven markers) is load-bearing only if the latent deviations are not redundant with those markers; the abstract supplies neither a correlation table between the latent score and the seven markers nor an ablation that removes the markers before adding the latent score.
[Abstract] Abstract: the held-out evaluation result assumes the graph autoencoder + ODE representation model was trained exclusively on the training partition with no information from the evaluation subset used for the downstream Cox model; the abstract does not state this separation explicitly, leaving open the possibility of circularity in the reported 0.785 figure.

minor comments (1)

[Abstract] The abstract refers to 'refitted pooled cohort equations' without specifying which covariates were included in the refit or how the refitting was performed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive review and for recognizing the technical contributions and the need for external validation. We respond point-by-point to the two major comments on the abstract.

read point-by-point responses

Referee: [Abstract] Abstract: the headline claim that the latent score adds incremental value (C-index rising from 0.704 to 0.785 versus 0.764 for the seven markers) is load-bearing only if the latent deviations are not redundant with those markers; the abstract supplies neither a correlation table between the latent score and the seven markers nor an ablation that removes the markers before adding the latent score.

Authors: We agree that explicit evidence of non-redundancy is required to support the incremental-value claim in the abstract. The full manuscript contains correlation analyses between the latent deviations and the seven markers together with an ablation that adds the latent score after the markers; these results are reported in the main text and supplementary material. Because of abstract length constraints the supporting details were omitted. We will revise the abstract to include a concise statement that the latent score supplies incremental information independent of the markers, with reference to the detailed results. revision: yes
Referee: [Abstract] Abstract: the held-out evaluation result assumes the graph autoencoder + ODE representation model was trained exclusively on the training partition with no information from the evaluation subset used for the downstream Cox model; the abstract does not state this separation explicitly, leaving open the possibility of circularity in the reported 0.785 figure.

Authors: We confirm that the graph autoencoder and heart-rate-aware neural ODE were trained exclusively on the training partition; the held-out subset was reserved solely for the downstream Cox regression and C-index evaluation. The abstract refers to a 'held-out evaluation subset' but does not explicitly state the training separation. To eliminate any ambiguity regarding possible leakage or circularity, we will revise the abstract to state clearly that the representation model was trained only on the training data. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper's chain proceeds from training a graph-mesh autoencoder plus heart-rate-aware neural ODE on cine CMR data to produce latent trajectories, defining a covariate-conditioned prior on the end-diastolic state, extracting deviations, and feeding those deviations into a Cox model whose performance is evaluated on an explicitly held-out subset. The reported C-index gain (0.704 to 0.785) is therefore measured on data unseen by the latent model, and the architecture itself supplies reconstruction and generative objectives independent of the downstream risk score. No equations, self-citations, or fitted parameters are shown to reduce the headline result to its own inputs by construction; the abstract itself flags the need for external validation, confirming the within-paper evaluation is treated as provisional rather than self-verifying.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities; the model necessarily contains many neural-network weights, ODE solver tolerances, and mesh-graph hyperparameters that are not enumerated here.

pith-pipeline@v0.9.1-grok · 5795 in / 1248 out tokens · 25696 ms · 2026-06-26T04:45:22.921351+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

12 extracted references · 10 canonical work pages

[1]

Cardiovascu- lar Imaging 15, 1563–1574

Understanding and improving risk assessment after myocardial infarctionusingautomatedleftventricularshapeanalysis. Cardiovascu- lar Imaging 15, 1563–1574. doi:10.1016/j.jcmg.2021.11.027. Dalca, A.V., Balakrishnan, G., Guttag, J., Sabuncu, M.R., 2019. Unsuper- visedlearningofprobabilisticdiffeomorphicregistrationforimagesand surfaces. Medical image analysi...

work page doi:10.1016/j.jcmg.2021.11.027 2021
[2]

Jama 323, 636–645

Predictiveaccuracyofapolygenicriskscore–enhancedprediction model vs a clinical risk score for coronary artery disease. Jama 323, 636–645. doi:10.1001/jama.2019.22241. Fonseca, C.G., Backhaus, M., Bluemke, D.A., Britten, R.D., Chung, J.D., Cowan, B.R., Dinov, I.D., Finn, J.P., Hunter, P.J., Kadish, A.H., et al.,

work page doi:10.1001/jama.2019.22241 2019
[3]

Bioinformatics 27, 2288–

Thecardiacatlasproject—animagingdatabaseforcomputational modeling and statistical atlases of the heart. Bioinformatics 27, 2288–
[4]

Frangi, A.F., Rueckert, D., Schnabel, J.A., Niessen, W.J., 2003

doi:10.1093/bioinformatics/btr360. Frangi, A.F., Rueckert, D., Schnabel, J.A., Niessen, W.J., 2003. Automatic constructionofmultiple-objectthree-dimensionalstatisticalshapemod- els: Application to cardiac modeling. IEEE transactions on medical imaging 21, 1151–1166. doi:10.1109/TMI.2002.804426. Fry, A., Littlejohns, T.J., Sudlow, C., Doherty, N., Adamska,...

work page doi:10.1093/bioinformatics/btr360 2003
[5]

Journal of the American College of Cardiology 79, e263–e421

2022 aha/acc/hfsa guideline for the management of heart failure: a report of the american college of cardiology/american heart associ- ation joint committee on clinical practice guidelines. Journal of the American College of Cardiology 79, e263–e421. doi:10.1161/CIR. 0000000000001063. Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H., ...

work page doi:10.1161/cir 2022
[6]

European Heart Journal 42, 3599–3726

2021 esc guidelines for the diagnosis and treatment of acute and chronic heart failure. European Heart Journal 42, 3599–3726. doi:10.1093/eurheartj/ehab368. Meng, Q., Bai, W., O’Regan, D.P., Rueckert, D., 2024. Deepmesh: Mesh- based cardiac motion tracking using deep learning. IEEE Transactions on Medical Imaging 43, 1489–1500. doi:10.1109/TMI.2023.334011...

work page doi:10.1093/eurheartj/ehab368 2021
[7]

Page 14 of 15 Latent ODE Modeling of Cine Cardiac MRI 25, 1374–1383

Evaluation of deep learning estimation of whole heart anatomy from automated cardiovascular magnetic resonance short-and long-axis analysesinukbiobank.EuropeanHeartJournal-CardiovascularImaging Brüggemann et al. Page 14 of 15 Latent ODE Modeling of Cine Cardiac MRI 25, 1374–1383. doi:10.1093/ehjci/jeae123. Ng, M., Guo, F., Biswas, L., Petersen, S.E., Piec...

work page doi:10.1093/ehjci/jeae123 2023
[8]

Sinclair,M.,Bai,W.,Puyol-Antón,E.,Oktay,O.,Rueckert,D.,King,A.P.,

doi:10.1038/s42256-025-01035-5. Sinclair,M.,Bai,W.,Puyol-Antón,E.,Oktay,O.,Rueckert,D.,King,A.P.,

work page doi:10.1038/s42256-025-01035-5
[9]

Fully automated segmentation-based respiratory motion correc- tion of multiplanar cardiac magnetic resonance images for large-scale datasets,in:InternationalConferenceonMedicalImageComputingand Computer-Assisted Intervention, pp. 332–340. Sohn, K., Lee, H., Yan, X., 2015. Learning structured output representa- tion using deep conditional generative models...

2015
[10]

Van Buuren, S., Groothuis-Oudshoorn, K., 2011

doi:10.1109/JBHI.2017.2652449. Van Buuren, S., Groothuis-Oudshoorn, K., 2011. mice: Multivariate imputation by chained equations in r. Journal of statistical software 45, 1–67. doi:10.18637/jss.v045.i03. Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N., Kaiser, Ł., Polosukhin, I., 2017. Attention is all you need, in: Advances in neural inf...

work page doi:10.1109/jbhi.2017.2652449 2017
[11]

Medical image analysis 13, 773–784

Modelling passive diastolic mechanics with quantitative mri of cardiac structure and function. Medical image analysis 13, 773–784. doi:10.1016/j.media.2009.07.006. World Health Organization, 2025. Cardiovascular diseases (cvds). URL: https://www.who.int/health-topics/cardiovascular-diseases. accessed 2026-04-22. Wu, Y., Jiahao, T.Z., Wang, J., Yushkevich,...

work page doi:10.1016/j.media.2009.07.006 2009
[12]

20804–20813

Nodeo:Aneuralordinarydifferentialequationbasedoptimization framework for deformable image registration, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 20804–20813. Xia,Y.,Chen,X.,Ravikumar,N.,Kelly,C.,Attar,R.,Aung,N.,Neubauer, S., Petersen, S.E., Frangi, A.F., 2022. Automatic 3d+t four-chamber cmr quantificatio...

work page doi:10.1016/j.media.2022.102498 2022

[1] [1]

Cardiovascu- lar Imaging 15, 1563–1574

Understanding and improving risk assessment after myocardial infarctionusingautomatedleftventricularshapeanalysis. Cardiovascu- lar Imaging 15, 1563–1574. doi:10.1016/j.jcmg.2021.11.027. Dalca, A.V., Balakrishnan, G., Guttag, J., Sabuncu, M.R., 2019. Unsuper- visedlearningofprobabilisticdiffeomorphicregistrationforimagesand surfaces. Medical image analysi...

work page doi:10.1016/j.jcmg.2021.11.027 2021

[2] [2]

Jama 323, 636–645

Predictiveaccuracyofapolygenicriskscore–enhancedprediction model vs a clinical risk score for coronary artery disease. Jama 323, 636–645. doi:10.1001/jama.2019.22241. Fonseca, C.G., Backhaus, M., Bluemke, D.A., Britten, R.D., Chung, J.D., Cowan, B.R., Dinov, I.D., Finn, J.P., Hunter, P.J., Kadish, A.H., et al.,

work page doi:10.1001/jama.2019.22241 2019

[3] [3]

Bioinformatics 27, 2288–

Thecardiacatlasproject—animagingdatabaseforcomputational modeling and statistical atlases of the heart. Bioinformatics 27, 2288–

[4] [4]

Frangi, A.F., Rueckert, D., Schnabel, J.A., Niessen, W.J., 2003

doi:10.1093/bioinformatics/btr360. Frangi, A.F., Rueckert, D., Schnabel, J.A., Niessen, W.J., 2003. Automatic constructionofmultiple-objectthree-dimensionalstatisticalshapemod- els: Application to cardiac modeling. IEEE transactions on medical imaging 21, 1151–1166. doi:10.1109/TMI.2002.804426. Fry, A., Littlejohns, T.J., Sudlow, C., Doherty, N., Adamska,...

work page doi:10.1093/bioinformatics/btr360 2003

[5] [5]

Journal of the American College of Cardiology 79, e263–e421

2022 aha/acc/hfsa guideline for the management of heart failure: a report of the american college of cardiology/american heart associ- ation joint committee on clinical practice guidelines. Journal of the American College of Cardiology 79, e263–e421. doi:10.1161/CIR. 0000000000001063. Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H., ...

work page doi:10.1161/cir 2022

[6] [6]

European Heart Journal 42, 3599–3726

2021 esc guidelines for the diagnosis and treatment of acute and chronic heart failure. European Heart Journal 42, 3599–3726. doi:10.1093/eurheartj/ehab368. Meng, Q., Bai, W., O’Regan, D.P., Rueckert, D., 2024. Deepmesh: Mesh- based cardiac motion tracking using deep learning. IEEE Transactions on Medical Imaging 43, 1489–1500. doi:10.1109/TMI.2023.334011...

work page doi:10.1093/eurheartj/ehab368 2021

[7] [7]

Page 14 of 15 Latent ODE Modeling of Cine Cardiac MRI 25, 1374–1383

Evaluation of deep learning estimation of whole heart anatomy from automated cardiovascular magnetic resonance short-and long-axis analysesinukbiobank.EuropeanHeartJournal-CardiovascularImaging Brüggemann et al. Page 14 of 15 Latent ODE Modeling of Cine Cardiac MRI 25, 1374–1383. doi:10.1093/ehjci/jeae123. Ng, M., Guo, F., Biswas, L., Petersen, S.E., Piec...

work page doi:10.1093/ehjci/jeae123 2023

[8] [8]

Sinclair,M.,Bai,W.,Puyol-Antón,E.,Oktay,O.,Rueckert,D.,King,A.P.,

doi:10.1038/s42256-025-01035-5. Sinclair,M.,Bai,W.,Puyol-Antón,E.,Oktay,O.,Rueckert,D.,King,A.P.,

work page doi:10.1038/s42256-025-01035-5

[9] [9]

Fully automated segmentation-based respiratory motion correc- tion of multiplanar cardiac magnetic resonance images for large-scale datasets,in:InternationalConferenceonMedicalImageComputingand Computer-Assisted Intervention, pp. 332–340. Sohn, K., Lee, H., Yan, X., 2015. Learning structured output representa- tion using deep conditional generative models...

2015

[10] [10]

Van Buuren, S., Groothuis-Oudshoorn, K., 2011

doi:10.1109/JBHI.2017.2652449. Van Buuren, S., Groothuis-Oudshoorn, K., 2011. mice: Multivariate imputation by chained equations in r. Journal of statistical software 45, 1–67. doi:10.18637/jss.v045.i03. Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N., Kaiser, Ł., Polosukhin, I., 2017. Attention is all you need, in: Advances in neural inf...

work page doi:10.1109/jbhi.2017.2652449 2017

[11] [11]

Medical image analysis 13, 773–784

Modelling passive diastolic mechanics with quantitative mri of cardiac structure and function. Medical image analysis 13, 773–784. doi:10.1016/j.media.2009.07.006. World Health Organization, 2025. Cardiovascular diseases (cvds). URL: https://www.who.int/health-topics/cardiovascular-diseases. accessed 2026-04-22. Wu, Y., Jiahao, T.Z., Wang, J., Yushkevich,...

work page doi:10.1016/j.media.2009.07.006 2009

[12] [12]

20804–20813

Nodeo:Aneuralordinarydifferentialequationbasedoptimization framework for deformable image registration, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 20804–20813. Xia,Y.,Chen,X.,Ravikumar,N.,Kelly,C.,Attar,R.,Aung,N.,Neubauer, S., Petersen, S.E., Frangi, A.F., 2022. Automatic 3d+t four-chamber cmr quantificatio...

work page doi:10.1016/j.media.2022.102498 2022