What question did this study set out to answer?

This research aims to develop a deep learning framework, CardioNets, to generate cardiac magnetic resonance (CMR) images from electrocardiograms (ECGs).

March 31, 2026NEJM AI

Generating Cardiac Magnetic Resonance Images from Electrocardiograms — A Multicenter Study

Q: What is the clinical evidence from this study?

Study design: Other. Population: Cardiovascular diseases (n=163586). Intervention: CardioNets (deep learning framework) vs. Baseline models, prior approaches, and human physicians. Primary outcome: Disease screening and phenotype estimation performance (including AUC, R2, SSIM, PSNR, and diagnostic accuracy).

View Full Paper

Ask AI

Bookmark

Key result

ECG-only CardioNets achieves ~14% higher diagnostic accuracy than physicians using both ECG and CMR.

n=163,586

Why the study?

CMR provides gold-standard structural and functional insights but is limited by cost and complexity, whereas ECG is accessible but lacks CMR granularity.

Does the CardioNets deep learning framework accurately generate CMR-level functional parameters and synthetic images from 12-lead ECG signals compared to baseline models and human physicians?

Population

159,819 training samples from five cohorts and independent clinical validation datasets of n=3,767

Comparison

CardioNets deep learning framework vs baseline models, prior approaches, and human physicians

Design

Multicenter model development and external validation study

Authors

ZDZhengyao DingZhejiang University of Science and Technology YHYujian HuFirst Affiliated Hospital Zhejiang University YXYouyao XuQuzhou City People's Hospital

Discussion

Member takes

Implication

May enable ECG-based estimation of CMR metrics in resource-limited settings; leaves open prospective validation before clinical use.

Key Points

This research aims to develop a deep learning framework, CardioNets, to generate cardiac magnetic resonance (CMR) images from electrocardiograms (ECGs).
Developed CardioNets framework for translating 12-lead ECG signals into CMR-level parameters and synthetic images.
Trained on 159,819 samples from five cohorts, including UK Biobank and MIMIC-IV-ECG, with external validation on 3,767 samples.
Employed cross-modal contrastive learning and generative pretraining to synthesize CMR images.
CardioNets improved cardiac phenotype regression R2 by 24.8% and cardiomyopathy AUC by up to 39.3% in the UK Biobank.
In MIMIC, detection of pulmonary hypertension increased AUC by 5.6%.
Generated CMR images showed a 36.6% higher SSIM and 8.7% higher PSNR compared to earlier methods.

Structured PICO

Does the CardioNets deep learning framework accurately generate CMR-level functional parameters and synthetic images from 12-lead ECG signals compared to baseline models and human physicians?

Population

163,586 samples from multiple cohorts, including the UK Biobank and MIMIC-IV-ECG, used to train and validate an ECG-to-CMR deep learning framework.

Exposure

CardioNets, a deep learning framework that translates 12-lead ECG signals into CMR-level functional parameters and synthetic images using cross-modal contrastive learning and generative pretraining.

Comparator

Baseline AI models, prior image generation approaches, and human physicians using both ECG and real CMR.

Outcome

Model performance across disease screening and phenotype estimation tasks (including cardiac phenotype regression R2, cardiomyopathy AUC, pulmonary hypertension detection AUC, image quality metrics SSIM and PSNR, and diagnostic accuracy).surrogate

CardioNets demonstrates that deep learning can synthesize CMR-level data and images from standard 12-lead ECGs, potentially offering a low-cost, scalable alternative for cardiovascular disease screening.

Cite This Study

Ding et al. (2026) studied Cardiovascular diseases (n=163,586). CardioNets (deep learning framework) vs. Baseline models, prior approaches, and human physicians was evaluated on Disease screening and phenotype estimation performance (including AUC, R2, SSIM, PSNR, and diagnostic accuracy). The CardioNets deep learning framework using only ECGs achieved 13.9% higher diagnostic accuracy in a reader study than human physicians using both ECG and real CMR.

synapsesocial.com/papers/6a025c6cedf6f48138594614 https://doi.org/10.1056/aioa2500549

View Full Paper

Ask AI

Bookmark

March 31, 2026NEJM AI

Generating Cardiac Magnetic Resonance Images from Electrocardiograms — A Multicenter Study

View Full Paper

Ask AI

Bookmark

Key result

ECG-only CardioNets achieves ~14% higher diagnostic accuracy than physicians using both ECG and CMR.

n=163,586

Why the study?

CMR provides gold-standard structural and functional insights but is limited by cost and complexity, whereas ECG is accessible but lacks CMR granularity.

Does the CardioNets deep learning framework accurately generate CMR-level functional parameters and synthetic images from 12-lead ECG signals compared to baseline models and human physicians?

Population

159,819 training samples from five cohorts and independent clinical validation datasets of n=3,767

Comparison

CardioNets deep learning framework vs baseline models, prior approaches, and human physicians

Design

Multicenter model development and external validation study

Authors

ZDZhengyao DingZhejiang University of Science and Technology YHYujian HuFirst Affiliated Hospital Zhejiang University YXYouyao XuQuzhou City People's Hospital

Discussion

Member takes

Implication

May enable ECG-based estimation of CMR metrics in resource-limited settings; leaves open prospective validation before clinical use.

Key Points

This research aims to develop a deep learning framework, CardioNets, to generate cardiac magnetic resonance (CMR) images from electrocardiograms (ECGs).
Developed CardioNets framework for translating 12-lead ECG signals into CMR-level parameters and synthetic images.
Trained on 159,819 samples from five cohorts, including UK Biobank and MIMIC-IV-ECG, with external validation on 3,767 samples.
Employed cross-modal contrastive learning and generative pretraining to synthesize CMR images.
CardioNets improved cardiac phenotype regression R2 by 24.8% and cardiomyopathy AUC by up to 39.3% in the UK Biobank.
In MIMIC, detection of pulmonary hypertension increased AUC by 5.6%.
Generated CMR images showed a 36.6% higher SSIM and 8.7% higher PSNR compared to earlier methods.

Structured PICO

Does the CardioNets deep learning framework accurately generate CMR-level functional parameters and synthetic images from 12-lead ECG signals compared to baseline models and human physicians?

Population

163,586 samples from multiple cohorts, including the UK Biobank and MIMIC-IV-ECG, used to train and validate an ECG-to-CMR deep learning framework.

Exposure

CardioNets, a deep learning framework that translates 12-lead ECG signals into CMR-level functional parameters and synthetic images using cross-modal contrastive learning and generative pretraining.

Comparator

Baseline AI models, prior image generation approaches, and human physicians using both ECG and real CMR.

Outcome

Cite This Study

synapsesocial.com/papers/6a025c6cedf6f48138594614 https://doi.org/10.1056/aioa2500549

Also Consider

Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context:

View Full Paper

Ask AI

Bookmark