PhD candidate McGill University Montréal, Quebec, Canada
Background: Myocardial edema is a hallmark of acute myocardial injury and plays a critical role in diagnosing conditions such as myocarditis. T2-weighted magnetic resonance imaging (MRI), which detects edema based on prolonged T2-decay times and T2-mapping, is the clinical gold standard. However, these techniques are limited by long acquisition times and protocol variability. Oxygen-sensitive cardiovascular magnetic resonance (OS-CMR), a contrast-free sequence, has shown potential to reflect both T1 and T2 tissue characteristics. While generative adversarial networks (GANs) have shown promise in synthesizing medical images, their application to T2-mapping remains limited. We hypothesized that OS-CMR encodes edema-relevant tissue information and developed a Residual GAN (R-GAN) with a clinically integrated pipeline to synthesize high-fidelity T2-parametric maps directly from OS-CMR, enabling edema detection without the need for dedicated T2 sequences.
METHODS AND RESULTS: We trained a R-GAN on 2,189 paired OS-CMR and T2-weighted images, following preprocessing steps including registration, and augmentation. The model’s performance was compared to Pix2Pix using peak signal-to-noise ratio (PSNR: 21.56–24.95 dB), structural similarity index (SSIM: 0.672–0.739), and Pearson correlation coefficient (PCC: 0.862–0.923) (Table 1). Signal intensity analysis between synthetic and original T2-maps showed close agreement in representative cases (e.g., edema ROI: 452 vs. 459; healthy ROI: 283 vs. 269). To assess the quality of synthesized T2-decay, we used Extended Phase Graph (EPG) simulations based on the synthetic image. The generated synthetic T2-curves closely matched the original, with Pearson and Spearman correlations of 1.0000, MAE ranging from 0.0007 to 0.0185, MSE from 0.0000 to 0.0004, RMSE from 0.0008 to 0.0207, R² values between 0.8712 and 0.9997, and SAD between 0.0029 and 0.0739. Additionally, T2 curve validation was performed in edema and healthy myocardial ROIs. In edema ROIs, correlations ranged from 0.99 to 1.0000, MAE from 0.0041 to 0.0286, MSE from 0.0000 to 0.0017, RMSE from 0.0041 to 0.0286, R² from 0.9995 to 1.0000, and SAD from 0.0163 to 0.1144. In healthy ROIs, Pearson and Spearman correlations ranged from 0.98 to 1.0000, MAE from 0.0063 to 0.0595, MSE from 0.0000 to 0.0168, RMSE from 0.0063 to 0.1503, R² from 0.9671 to 1.0000, and SAD from 0.0253 to 0.5668. These results confirm that R-GAN can reliably synthesize high-quality T2 parametric maps (Figure 1).
Conclusion: Our R-GAN successfully synthesizes T2-parametric maps from OS-CMR images, enabling myocardial edema detection without additional sequences. This approach highlights the potential to transform functional MRI into a tool for structural tissue characterization, offering a time-efficient, non-invasive alternative for cardiac diagnostics.
