The novel R-GAN synthesized T2 maps from OS-CMR with up to 24.95 dB PSNR, 0.739 SSIM, and 0.923 PCC, achieving 99% correlation with original edema T2 curves.
Does a novel R-GAN framework improve the synthesis of T2 parametric maps from OS-CMR images compared to the Pix2Pix model for myocardial edema detection?
A novel R-GAN framework can accurately synthesize T2 parametric maps from OS-CMR images, offering a promising non-invasive approach for myocardial edema detection.
Abstract Background Myocardial edema is a key factor of acute myocardial injury that substantially impairs microvascular function 1. Although T2-weighted MRI is essential for identifying this edema, its performance is often compromised by motion artifacts, prolonged acquisition times, and heterogeneous imaging protocols. 2, 3 In contrast, needle-free OS-CMR promising approach to provides functional insights and included T2 information 4-6. With recent advances in deep learning, especially in generative models 7, the synthesis of medical images has become a promising avenue. This study introduces a novel R-GAN coupled with an edema validation pipeline designed to synthesize T2 parametric maps from OS-CMR, thereby facilitating accurate edema detection. Purpose The purpose of this study is to develop and validate a novel R-GAN framework for synthesizing T2 parametric maps from OS-CMR images, enabling precise detection of myocardial edema. Methods Two independent datasets comprising matched OS-CMR and T2-weighted images from 2189 patients were analyzed. Preprocessing steps included normalization, data augmentation, and image registration. The proposed R-GAN (figure1), which incorporates residual blocks, was benchmarked against the Pix2Pix model using performance metrics such as PSNR, SSIM, and PCC. The evaluation focused on both edematous and healthy myocardial regions through signal intensity measurements, Extended Phase Graph (EPG) simulations, and T2 curve analyses performed with Python scripts and CVI42 software (Figure2). Results The R-GAN consistently outperformed the Pix2Pix model across all evaluated metrics. It achieved PSNR values ranging from 21.56 to 24.95 dB and SSIM scores between 0.672 and 0.739, and PCC scores between 0.862 to 0.923 compared to Pix2Pix’s 18.80–19.20 dB and 0.055–0.659 and 0.802 – 0.890, respectively. The edema signal intensity remained robust, with the original values of 452 for the edema ROI and 283 for the healthy ROI closely matching the synthesized values of 459.0 and 269.0, respectively. Extended Phase Graph (EPG) simulations revealed a 99% correlation between the synthesized T2 curves and the original data. Furthermore, analysis of two regions of interest, representing edematous and healthy myocardium, validated the fidelity of the synthetic T2 curves, as demonstrated by error metrics showing MAE values from 0.0007 to 0.0185, RMSE as low as 0.0008, and consistently high R² values (≥0.87). Notably, T2 curves for both edematous and healthy regions exhibited an almost perfect correlation (0.99) between the synthesized and ground truth data. Conclusion The novel R-GAN method demonstrates the feasibility of generating high-quality T2 maps from OS-CMR images, offering a promising non-invasive approach for myocardial edema detection. Further studies are warranted to validate these findings.Figure1 Figure2
lotfikazemi et al. (2025) studied this question. The novel R-GAN synthesized T2 maps from OS-CMR with up to 24.95 dB PSNR, 0.739 SSIM, and 0.923 PCC, achieving 99% correlation with original edema T2 curves.