[Paper Review] Chain of Flow: A Foundational Generative Framework for ECG-to-4D Cardiac Digital Twins

Haofan Wu, Nay Aung+4 more|arXiv (Cornell University)|Feb 26, 2026

Congenital heart defects research0 citations

TL;DR

Chain of Flow (COF) is an ECG-conditioned generative framework that reconstructs a full 4D cardiac structure and motion from a single cardiac cycle, enabling patient-specific digital twins and downstream cardiac simulations.

ABSTRACT

A clinically actionable Cardiac Digital Twin (CDT) should reconstruct individualised cardiac anatomy and physiology, update its internal state from multimodal signals, and enable a broad range of downstream simulations beyond isolated tasks. However, existing CDT frameworks remain limited to task-specific predictors rather than building a patient-specific, manipulable virtual heart. In this work, we introduce Chain of Flow (COF), a foundational ECG-driven generative framework that reconstructs full 4D cardiac structure and motion from a single cardiac cycle. The method integrates cine-CMR and 12-lead ECG during training to learn a unified representation of cardiac geometry, electrophysiology, and motion dynamics. We evaluate Chain of Flow on diverse cohorts and demonstrate accurate recovery of cardiac anatomy, chamber-wise function, and dynamic motion patterns. The reconstructed 4D hearts further support downstream CDT tasks such as volumetry, regional function analysis, and virtual cine synthesis. By enabling full 4D organ reconstruction directly from ECG, COF transforms cardiac digital twins from narrow predictive models into fully generative, patient-specific virtual hearts. Code will be released after review.

Motivation & Objective

• Motivate the need for a fully generative, patient-specific cardiac digital twin that integrates anatomy, physiology, and motion.
• Develop an ECG-driven model that reconstructs 4D cine-CMR volumes from ECG without relying on repeated CMR imaging.
• Leverage multimodal training (cine-CMR + ECG) to learn a unified representation of geometry, electrophysiology, and motion.
• Enable downstream CDT tasks such as volumetry, regional function analysis, and virtual cine synthesis from the generated 4D heart.

Proposed method

• Introduce TOPPR for topology-preserving registration to estimate time-coherent 3D displacement fields between CMR volumes.
• Compute discrete deformation samples to parameterize an underlying spatiotemporal velocity field over the cardiac cycle.
• Learn an ECG-conditioned dynamic flow v_theta(x,t,c) that ties ECG-based electrophysiology signals to subject-specific anatomy via flow matching.
• Supervise motion learning with a reference velocity field derived from registered deformation trajectories.
• Infer a subject-specific 4D heart by integrating the velocity field via an ODE solver and applying the resulting deformation to the reference anatomy.

Experimental results

Research questions

• RQ1 Can ECG data alone drive accurate, fully 4D cardiac reconstruction that preserves anatomical fidelity and realistic motion?
• RQ2 Does the ECG-conditioned generative framework generalize across slice positions, resolutions, and diverse ECG-derived disease categories?
• RQ3 Can the generated 4D hearts support clinically relevant indices and downstream cardiac digital twin tasks (volumetry, regional function, virtual cine)?
• RQ4 What is the contribution of each component (anatomical anchoring, ECG-driven dynamics, and segmentation-consistency supervision) to reconstruction quality?
• RQ5 Is the approach robust to variations in image resolution and acquisition settings typical of large cohorts?

Key findings

• COF achieves state-of-the-art performance across image-level metrics on UK Biobank data (SSIM 0.984, PSNR 28.46, FID 6.39, FVD 17.60, M-Corr. 0.474, M-SSIM 0.894).
• COF outperforms ablated variants and other baselines in both image fidelity and segmentation accuracy (LV Dice 0.87, LV IoU 0.80; RV Dice 0.74, RV IoU 0.61; Myo Dice 0.85, Myo IoU 0.74).
• Segmentation-consistency supervision and TOPPR are critical for preserving anatomical fidelity and realistic motion; removing TOPPR or segmentation loss degrades multiple metrics.
• COF maintains stable performance across slice positions and resolutions, showing robust multi-slice and multi-resolution representation.
• Population-level analyses show COF preserves clinically meaningful 4D function (EDV, ESV, SV, EF, CO) across diverse ECG-derived categories and yields correlated LV volume–time curves with real data.
• Qualitative case studies (e.g., hypertrophy and ischemia/infarction) illustrate anatomically plausible and functionally consistent generated hearts.