Chain of Flow: ECG-Conditioned 4D Cardiac Cine Generation from Patient-Specific Anatomical Anchor

Haofan Wu; Nay Aung; Theodoros N. Arvanitis; Joao A. C. Lima; Steffen E. Petersen; Le Zhang

cs.CVFeb 26, 2026

Chain of Flow: ECG-Conditioned 4D Cardiac Cine Generation from Patient-Specific Anatomical Anchor

Authors: Haofan Wu, Nay Aung, Theodoros N. Arvanitis, Joao A. C. Lima, Steffen E. Petersen, Le Zhang

Organizations: School of Engineering, College of Engineering and Physical Sciences, University of Birmingham, Birmingham, United Kingdom · William Harvey Research Institute, NIHR Barts Biomedical Research Centre, Queen Mary University London, London, United Kingdom · Barts Heart Centre, St Bartholomew’s Hospital, Barts Health NHS Trust, London, United Kingdom · Division of Cardiology, Johns Hopkins University School of Medicine, Baltimore, Maryland, United States

Abstract

Cardiac cine magnetic resonance imaging (MRI) is central to functional cardiac assessment, yet a full current cine sequence may not always be directly available at the point of analysis. We introduce Chain of Flow (COF), an electrocardiography (ECG)-conditioned framework that combines patient-specific MRI and current ECG for subject-specific 4D cardiac cine generation. On the UK Biobank dataset, COF achieves strong image-level fidelity and downstream function-oriented performance on a shared same-visit evaluable benchmark. Multi-slice and multi-resolution analyses indicate stable structural generation quality across the short-axis stack and heterogeneous acquisition resolutions. Controlled phase-robustness analyses across resampled input MRI phases further provide same-visit proxy support for patient-specific MRI plus current ECG when a target MRI phase is not directly observed. A cross-visit route provides exploratory serial evidence, with the clearest gains in current-facing region-of-interest readout. Disease-category functional audits, case-level volume-trajectory evidence review further delineate where the current patient-specific MRI plus ECG formulation remains stable for anatomy-aware downstream cardiac analysis. Code is available at https://anonymous.4open.science/r/COF-paper-release-C88B.

Figures & tables

Fig. 1: Overview of COF for ECG-guided 4D cardiac volume generation from patient-specific MRI and current ECG. The framework integrates CMR volumes and 12-lead ECG during training to learn a unified representation of cardiac anatomy, electrophysiology, and motion dynamics. Cine-CMR provides anatomical geometry and motion supervision, while ECG supplies subject-specific electrophysiological dynamics over a single cardiac cycle. These multimodal signals are fused within the proposed Chain of Flow, which couples geometry encoding, electro-physiological flow, and motion propagation to form an ECG-conditioned dynamic representation. At inference, the learned model takes patient-specific MRI together with current ECG to generate a subject-specific 4D cine sequence with individualised cardiac motion. The resulting representation supports downstream analyses including chamber volumetry, regional functional assessment, and anatomy-aware dynamic cardiac analysis.

Fig. 2: Algorithmic pipeline of registration-guided motion modelling and ECG-conditioned cine generation. The pipeline starts with a volumetric registration teacher that estimates 3D displacement fields between source and target CMR volumes at different cardiac phases. Discrete deformation samples obtained from registration are parameterised as a continuous spatiotemporal velocity field. A segmentation-guided teacher supervision hub constrains the registration using anatomy-aware Dice loss. The resulting subject-specific motion dynamics are used as supervision for flow matching, linking ECG signals to physiological motion priors. Finally, the learned velocity field is integrated over time via ODE-based integration to produce a 4D cine sequence with patient-specific cardiac motion.

Protocol stage	Cohort definition	$N$	Use in study
Generation entrance cohort	Paired UKB ECG–CMR cohort	10,000	generation route
Shared benchmark	Final same-visit benchmark	426	main results
Evaluator source	Annotated 4D source cohort	600	assessor source
Overlap removed	Subject overlap removed	9	independence audit
Evaluator-training cohort	Subject-disjoint assessor cohort	591	evaluator route
Cross-visit serial cohort	Dual-visit serial cohort	62	serial analysis

Table 1: Protocol transparency summary for the cohorts used in this manuscript. The table records the auditable routing from model development to the shared evaluable benchmark, evaluator training cohort, and cross-visit serial cohort.

Method	FVD $\!\downarrow$	FID $\!\downarrow$	PSNR $\!\uparrow$	SSIM $\!\uparrow$
COF	12.92 $\pm$ 0.5	4.54 $\pm$ 0.1	28.74 $\pm$ 0.85	0.985 $\pm$ 0.004
COF-w/Diff.	31.60 $\pm$ 0.7	8.82 $\pm$ 0.2	26.86 $\pm$ 0.77	0.978 $\pm$ 0.005
X-Dyna	19.38 $\pm$ 0.6	7.28 $\pm$ 0.1	27.12 $\pm$ 0.84	0.984 $\pm$ 0.004
MoFA-Video	47.24 $\pm$ 1.5	19.48 $\pm$ 0.4	27.01 $\pm$ 0.80	0.982 $\pm$ 0.004
EchoDiffusion	98.00 $\pm$ 1.8	66.48 $\pm$ 0.9	28.53 $\pm$ 0.73	0.986 $\pm$ 0.004
ECHOPulse	608.24 $\pm$ 7.7	437.35 $\pm$ 3.0	11.87 $\pm$ 0.63	0.361 $\pm$ 0.066

Table 2: Primary image-level comparison on the shared evaluable benchmark. Lower is better for FVD/FID. FVD and FID scores are reported as mean

\pm

spread from 50 resampled evaluations on aligned 426-subject DINOv2 feature bundles Oquab et al. [2023] , while PSNR and SSIM are reported as mean

\pm

standard deviation across subjects. Here, FVD and FID denote DINOv2-based Fréchet-style temporal and image distribution distances, rather than standard Inflated 3D or Inception-feature implementations.

Fig. 3: ECG-conditioned 4D cardiac cine generation results. The model conditions on a full 12-lead electrocardiogram, while the top trace shows one displayed lead from the aligned cardiac cycle for visual reference; the R-wave and T-wave markers indicate the temporal landmarks used to orient the displayed ED/ES phases within the cine sequence. The figure also shows the short-axis slice-stack context and the generated multi-slice cine examples across selected cardiac phases. For each phase, the four rows show the reference CMR slices, the generated slices, the reference labels from the real cine sequence, and the labels obtained from the generated slices through the automatic anatomical evaluator.

Method	ESV MAE $\!\downarrow$	EF MAE $\!\downarrow$	SV MAE $\!\downarrow$	Vol. curve RMSE $\!\downarrow$	ES phase err. $\!\downarrow$
COF	16.77	0.123	17.36	14.40	2.31
COF-w/Diff.	62.07	0.476	64.97	34.92	6.95
X-Dyna	35.57	0.236	37.04	22.47	2.05
MoFA-Video	32.18	0.217	33.32	19.67	2.07
EchoDiffusion	33.17	0.531	76.42	33.38	14.70
ECHOPulse	46.39	0.997	56.58	76.66	9.24

Table 3: Primary downstream function comparison under the independent automatic evaluator. Lower is better for all five reported metrics.

Fig. 4: (a) Image-level tradeoff view aligned with the main comparison in Table 2 after direction-aware normalisation. Larger area indicates a better overall image-level balance across the four reported metrics. (b) Function-first comparison aligned with the main comparison in Table 3 after the same direction-aware normalisation to summarise the downstream ranking.

Fig. 5: Per-slice anatomical agreement across cardiac phases. Slice-wise metrics are reported at end-diastole and end-systole after processing both real and generated cine with the independent automatic anatomical assessor. Slices are ordered from basal to apical (ranks 2–7) and evaluated separately for the left ventricle, right ventricle, and myocardium. Dice and IoU heatmaps (a, b, e, f) use a normalised 0–1 scale, whereas HD95 heatmaps (c, d) report 2D per-slice surface distances computed with in-plane spacing. Each cell is the subject-mean value for the corresponding structure and slice rank.

Fig. 6: Resolution-wise robustness. Dice (top), IoU (middle), and HD95 (bottom) are plotted against the in-plane pixel spacing $s_{x}$ (mm); points denote individual subjects, solid lines linear trend fits, and outlined markers spacing-bin means for LV, RV, and myocardium.

Frame	FVD	Corr.	SSIM	Macro	LV	EF
0	15.10	0.52	0.91	0.77	0.84	0.31
1	13.40	0.49	0.91	0.75	0.83	0.48
2	13.29	0.45	0.91	0.75	0.82	0.54
3	13.99	0.47	0.91	0.77	0.84	0.34
4	13.92	0.47	0.91	0.77	0.84	0.34
5	14.24	0.43	0.91	0.75	0.82	0.60

Table 4: Reference-anchor phase analysis on the shared evaluable benchmark. Subject and visit are fixed while MR and ECG are shifted together across 11 canonical anchor frames; frame 0 is the default inference anchor. Columns report FVD, motion correlation (Corr.), motion SSIM, macro Dice, LV Dice, and EF MAE.

Method	LV Dice	Myo Dice	Macro Dice	Prior curve Corr.	Prior EF drift	Prior SV drift
Prior	0.583	0.552	0.464	1.000	0.000	0.000
Prior+ECG	0.571	0.545	0.452	0.854	0.080	10.592
Reg.prior+ECG	0.731	0.640	0.607	0.908	0.083	11.234
Current+ECG	0.894	0.858	0.832	0.950	0.067	10.346

Table 5: Cross-visit serial support on the 62-subject same-subject dual-visit cohort. Prior denotes the carry-forward baseline, Prior+ECG the unregistered serial branch, Reg.prior+ECG the registered serial update, and Current+ECG the same-visit upper-bound context. The first three columns report current-facing ROI Dice, followed by prior-trajectory curve correlation and EF/SV drift. The last three columns report prior-trajectory stability rather than current-visit functional accuracy.

Fig. 7: Representative case-level visual support for the cross-visit serial ROI analysis. Panels show the current MRI, the prior MRI, the registered update, and the ROI contour comparison. In the ROI panel, solid cyan/light-blue contours denote the ground-truth LV/myocardium, and dashed red/orange contours denote the registered prediction.

Fig. 8: Representative cross-visit registration audit after prior-to-current frame 0 registration. Rows show raw prior, registered prior, current reference, and raw/registered overlays; red highlights indicate residual non-overlap.

Metric	ROI	non-ROI
MAE (raw)	0.1955	0.2823
MAE (reg.)	0.1293	0.2574
MAE drop	0.0662	0.0249
SSIM (raw)	0.6197	0.7345
SSIM (reg.)	0.7668	0.7572
SSIM gain	0.1472	0.0227

Table 6: ROI-versus-non-ROI decomposition of prior-to-current frame 0 registration on the 62-subject dual-visit cohort. ROI denotes the current-anchor cardiac union mask (LV+RV+Myo), and non-ROI denotes the body foreground minus ROI. Post-registration residual share denotes the fraction of total post-registration absolute error contained within ROI versus non-ROI.

Fig. 9: Population-level bootstrap distributions of the selected downstream functional readouts across ECG-derived disease categories on the shared evaluable benchmark.

Fig. 10: Representative disease cases showing ED/ES geometry and LV volume–time curves for Hypertrophy and Ischaemia / Infarction.

Setting

Modules

Image metrics

Seg. Dice / IoU

REA

ECG

TOPPR

SSIM

\!\uparrow

PSNR

\!\uparrow

FID

\!\downarrow

FVD

\!\downarrow

M - Corr.

\uparrow

M - SSIM

\uparrow

Myo

Module ablations

COF module setting

✓

0.986

29.21

7.98

18.58

0.548

0.905

N/A

w/o REA

\times

✓

0.979

26.79

7.53

18.91

0.226

0.865

N/A

w/o Temp. ECG

✓

\times

✓

0.978

26.76

14.50

25.24

0.512

0.896

N/A

w/o TOPPR

✓

\times

0.715

15.26

503.87

637.26

0.072

0.285

N/A

Table 7: Ablation study of COF on the UK Biobank dataset. Upper: module-wise ablations. Lower: segmentation-consistency-loss ablations from a separate internal sweep, intended for within-block trend reading rather than direct protocol-identical comparison to the reported main-table COF result. ‘No max field’ removes the maximum-displacement-field regularisation used in this sweep.

Medical Image Synthesis

Medical image synthesis creates, translates, reconstructs, and edits clinical images to fill missing data and support analysis.

See all 310 papers

One Frame, Full Heartbeat: ECG-Free Cardiac Cine MRI Synthesis via Phase-Conditioned Flow Matching
Oct 7, 2026Shiyi Wang, Ruochen Sun, Peirong Liu +2Medical Image Synthesis Cardiac MRI
Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD, USA · Gordon Center for Medical Imaging, Department of Radiology, Massachusetts General Hospital and Harvard Medical School, Boston, MA, USA · Center for Advanced Medical Computing and Analysis, Department of Radiology, Massachusetts General Hospital and Harvard Medical School, Boston, MA, USA
One Frame, Full Heartbeat: ECG-Free 4D Cardiac Cine MRI Synthesis via Radial-Decomposed Flow Matching
Oct 6, 2026Shiyi Wang, Ruochen Sun, Xiang Li +2Medical Image Synthesis Cardiac MRI
Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD, USA · Gordon Center for Medical Imaging, Department of Radiology, Massachusetts General Hospital and Harvard Medical School, Boston, MA, USA · Center for Advanced Medical Computing and Analysis, Department of Radiology, Massachusetts General Hospital and Harvard Medical School, Boston, MA, USA
ECGFlowCMR: Pretraining with ECG-Generated Cine CMR Helps Cardiac Disease Classification and Phenotype Prediction
Jan 28, 2026Xiaocheng Fang, Zhengyao Ding, Guangkun Nie +9Synthetic Data Pretraining Medical Diagnosis
State Key Laboratory of General Artificial Intelligence, Peking University · School of Intelligence Science and Technology, Peking University, Beijing, China · Polytechnic Institute of Zhejiang University, Zhejiang University, Hangzhou, China +4

Oct 7, 2026J/K move · Enter open · S save

One Frame, Full Heartbeat: ECG-Free Cardiac Cine MRI Synthesis via Phase-Conditioned Flow Matching

Shiyi Wang, Ruochen Sun, Peirong Liu, Xiang Li+1

Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD, USA · Gordon Center for Medical Imaging, Department of Radiology, Massachusetts General Hospital and Harvard Medical School, Boston, MA, USA · Center for Advanced Medical Computing and Analysis, Department of Radiology, Massachusetts General Hospital and Harvard Medical School, Boston, MA, USA

Medical Image Synthesis Cardiac MRI Flow Matching

Open paper AskPDF arXiv

Cine cardiovascular magnetic resonance (CMR) analysis relies on multi-frame sequences capturing the full cardiac cycle. However, standard multi-frame acquisition depends heavily on electrocardiogram (ECG) gating and repeated breath-holds, posing challenges in uncooperative populations, resource-limited settings, and temporally corrupted datasets. Existing methods that synthesize full cardiac sequences either rely on explicit ECG signals to parameterize myocardium function, or employ deformable registration without physiological constraints, failing to faithfully reproduce clinically relevant dynamic metrics such as ejection fraction (EF) and ventricular contraction magnitude. We present PhaseFlow, a unified generative framework that overcomes both limitations. PhaseFlow estimates a non-linear cardiac phase signal directly from the input sequence via a segmentation-derived left-ventricular (LV) area curve, capturing the asymmetric dynamics of systole and diastole without any ECG dependency. At inference, this phase signal is provided by a pathology-specific template, informing phase-specific frame generation. A rectified flow model conditioned on the phase and slice position synthesizes the full cardiac motion trajectory in the latent space, decoded into a diffeomorphic displacement field that warps end-diastole pixel intensities directly, eliminating the reconstruction blur often accompanying the variational autoencoder. On the ACDC benchmark, PhaseFlow achieves superior physiological fidelity and image realism, with best LV volume curve

R^2

, structural similarity (SSIM) and generative quality (FID) among all baselines. Ablation studies confirm that each proposed component contributes measurably to the overall performance.

Abstract

Medical Image Synthesis

Explore similar work

One Frame, Full Heartbeat: ECG-Free Cardiac Cine MRI Synthesis via Phase-Conditioned Flow Matching

One Frame, Full Heartbeat: ECG-Free 4D Cardiac Cine MRI Synthesis via Radial-Decomposed Flow Matching

ECGFlowCMR: Pretraining with ECG-Generated Cine CMR Helps Cardiac Disease Classification and Phenotype Prediction