PhD candidate McGill University Montréal, Quebec, Canada
Background: Myocardial fibrosis is a key indicator in various cardiomyopathies and is traditionally assessed using native T1-mapping and late gadolinium enhancement (LGE). However, contrast agents pose limitations for frequent or repeat imaging and may be contraindicated in some patients. We propose a Residual Generative Adversarial Network (R-GAN) that synthesizes native T1 parametric maps directly from non-contrast oxygenation-sensitive cardiovascular magnetic resonance (OS-CMR) images, enabling contrast-free fibrosis detection by extracting T1 and LGE-like patterns embedded within OS-CMR.
METHODS AND RESULTS: A total of 2044 paired OS-CMR and native T1-maps were used to train the R-GAN model with and without data augmentation. Preprocessing steps included image registration and normalization. Compared to the Pix2Pix baseline, the augmented R-GAN demonstrated superior performance with a structural similarity index (SSIM) of 0.810, peak signal-to-noise ratio (PSNR) of 16.01, and Pearson correlation coefficient (PCC) of 0.748. Synthesized T1 maps and signal intensity patterns were assessed using CVI42 across ischemic cardiomyopathy (ICMP), non-ischemic cardiomyopathy (NICMP), and amyloidosis cases. ROIs were manually selected in fibrotic and healthy myocardium, and T1 recovery curves were extracted from both synthesized and original maps. Correlation analysis of T1 curves showed strong alignment between synthesized and original signals: Case 1 yielded a Pearson correlation of 0.9815 and a Spearman correlation of 0.9985; Case 2 achieved perfect correlations (1.0000 for both); and Case 3 showed a Pearson correlation of 0.9717 and a Spearman correlation of 0.9984. To evaluate the biophysical validity of the generated maps, Extended Phase Graph (EPG) simulations were performed. In these simulations, the synthesized T1 curves closely matched their original counterparts, with mean squared error (MSE) values of 0.0001, 0.0002, and 0.0010; root mean squared error (RMSE) values of 0.0105, 0.0125, and 0.0323; and mean absolute error (MAE) values of 0.0054, 0.0069, and 0.0160 for Cases 1 through 3, respectively. The corresponding R² values were 0.9965, 0.9958, and 0.9078. Pearson correlations for EPG curves were 0.9999, 0.9999, and 0.9981, and Spearman correlations were 1.000 in all three cases. These results confirm a high degree of consistency in signal behavior between synthetic and original T1 data (see Figure 1).
Conclusion: The proposed R-GAN framework accurately generates native T1 maps from OS-CMR images without the need for contrast agents. Augmentation substantially improved performance. The model demonstrates strong generalizability and clinical validity, with visual and quantitative evaluations, both in CVI42 and through EPG simulation, highlighting its potential for non-invasive myocardial fibrosis assessment.
