Generating Cardiac Magnetic Resonance Images from Electrocardiograms - A Multicenter Study.
Authors
Affiliations (9)
Affiliations (9)
- College of Computer Science and Technology, Zhejiang University, Hangzhou, Zhejiang Province, China.
- Department of Vascular Surgery, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, Zhejiang Province, China.
- Department of Vascular Surgery, Quzhou People's Hospital, Quzhou, Zhejiang Province, China.
- Department of Cardiology, The Second Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, Zhejiang Province, China.
- Department of Radiology, The Second Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, Zhejiang Province, China.
- Department of Cardiology, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, Zhejiang Province, China.
- Department of Cardiology, Huzhou Central Hospital, Huzhou, Zhejiang Province, China.
- China Ship Scientific Research Center, Wuxi, Jiangsu Province, China.
- Guangdong Transtek Medical Electronics, Zhongshan, Guangdong Province, China.
Abstract
Cardiovascular disease (CVD) requires early and accurate diagnosis to improve outcomes. Cardiac magnetic resonance (CMR) imaging provides gold-standard functional and structural insights but remains limited by accessibility and complexity. Electrocardiogram (ECG) is widely available but lacks CMR's granularity. We propose CardioNets, a deep learning framework that translates 12-lead ECG into CMR-aligned functional parameters and synthetic images, enabling scalable, low-cost cardiac assessment. CardioNets aligns the ECG with CMR-derived latent representations via crossmodal contrastive learning, then generates CMR images from the ECG using masked autoregressive modeling. The study utilized 159,819 samples from the U.K. Biobank (UKB), the Medical Information Mart for Intensive Care (MIMIC) IV ECG database, and two external clinical datasets for model development and evaluation. The performance in cardiac measurement regression and disease detection was evaluated against ECG-only and CMR-based baseline models, and the synthesized CMR image quality was compared with state-of-the-art methods. A reader study compared CardioNets with physicians using ECG and clinically acquired CMR images. In the UKB dataset, CardioNets improved cardiac measurement regression (R<sup>2</sup>=0.310; 95% confidence interval [CI], 0.304 to 0.315) compared with the best ECG baseline (R<sup>2</sup>=0.269; 95% CI, 0.265 to 0.272; P<0.0001). For cardiomyopathy detection, CardioNets achieved an area under the receiver operator curve (AUROC) of 0.890 (95% CI, 0.836 to 0.944), outperforming the best ECG baseline (AUROC=0.867; 95% CI, 0.812 to 0.921; P<0.01), and demonstrating performance comparable to a CMR-based model (AUROC=0.906; 95% CI, 0.890 to 0.922; P=0.514). In MIMIC-IV for pulmonary hypertension detection, CardioNets (AUROC=0.879; 95% CI, 0.853 to 0.903) outperformed the best ECG baseline (AUROC=0.853; 95% CI, 0.824 to 0.881; P<0.001). Synthesized CMR images by CardioNets achieved a structural similarity index measure of 0.205 (95% CI, 0.202 to 0.208), outperforming the state-of-the-art method (0.122; 95% CI, 0.120 to 0.125; P<0.0001). In the reader study, CardioNets achieved an accuracy of 0.874 (95% CI, 0.800 to 0.923), outperforming participating readers. CardioNets translates ECGs into CMR-level insights, improving CVD detection performance and enabling scalable accessibility. Prospective studies are warranted to validate clinical deployment. (Funded by the National Key Research and Development Program of China and others.).