An AI framework integrating DeepOxyMap and a Residual Generative Adversarial Network generated synthetic T1 maps directly from contrast-free oxygenation-sensitive CMR that demonstrated excellent agreement with native T1 values (r = 0.976, p < 0.001).
Does an AI framework integrating DeepOxyMap and R-GAN accurately generate synthetic T1 maps from contrast-free oxygenation-sensitive CMR in patients with hypertrophic cardiomyopathy and cardiac amyloidosis?
An AI framework can generate quantitative synthetic T1 maps from contrast-free oxygenation-sensitive CMR, offering a rapid, gadolinium-free alternative for myocardial tissue characterization in hypertrophic cardiomyopathy and cardiac amyloidosis.
Effect estimate: r = 0.976
p-value: p=<0.001
Background Cardiovascular magnetic resonance (CMR) techniques provide detailed myocardial tissue characterization. However, LGE requires the administration of contrast agents, while T1/T2 mapping involves prolonged acquisition times, sensitivity to motion artifacts, and protocol complexity. LGE can also be limited in differentiating cardiac amyloidosis from hypertrophic cardiomyopathy (HCM) due to overlapping enhancement patterns. Oxygenation-sensitive CMR (OS-CMR), offers a rapid, contrast-free alternative but lacks direct quantitative outputs. Methods We propose a unified AI framework integrating DeepOxyMap and a Residual Generative Adversarial Network (R-GAN) to extract latent features and generate synthetic T1 maps from OS-CMR. DeepOxyMap was trained on 117 OS-CMR images using a VGG19-based architecture to identify fibrosis-related myocardial patterns via heatmaps. The R-GAN was subsequently trained on 2,044 OS-CMR images from two independent cohorts to generate T1-parametric maps directly from OS-CMR. Performance was assessed using image similarity metrics, ROI-based clinical validation, T1 recovery curve analysis, and Extended Phase Graph simulations. Results DeepOxyMap achieved an accuracy of of 83.5%, with a precision of 85.5% and a recall of 79.7%.AUC values were 0.93 (ICMP), 0.88 ( N ICMP), and 0.97 (healthy). Heatmaps showed strong spatial correspondence with fibrosis regions on LGE images. In the second stage, the R-GAN outperformed Pix2Pix (SSIM: 0.810, PSNR: 16.01, PCC: 0.748). EPG-based validation demonstrated near-perfect agreement between synthetic and theoretical T1 recovery curves ( R 2 0.9994). T1 curve agreement was also strong in ROI-based analysis. HCM: MAE = 0.0970, MSE = 0.0106, RMSE = 0.1030, R 2 = 0.7764, Pearson = 0.9658, Spearman = 0.9286. Amyloidosis: MAE = 0.0988, MSE = 0.0141, RMSE = 0.1187, R 2 = 0.8599, Pearson = 0.9718, Spearman = 0.9762. Synthetic T1 maps preserved disease-specific signal patterns. Synthetic and native T1 values demonstrated excellent agreement ( r = 0.976, p 0.001) with minimal bias (−6.7 ms). Conclusion This study demonstrates that OS-CMR contains latent tissue information that can be leveraged to generate quantitative T1 maps using AI. The proposed contrast-free framework enables rapid acquisition, eliminates the need for gadolinium, and preserves clinically relevant myocardial patterns, offering a promising approach for non-invasive cardiac tissue characterization in clinical practice.
Lotfikazemi et al. (2026) studied Hypertrophic cardiomyopathy and cardiac amyloidosis (n=2,143). Synthetic T1 mapping using DeepOxyMap and R-GAN from OS-CMR vs. Native T1 mapping (ground truth) was evaluated on Agreement between synthetic and native T1 values (r = 0.976, p=<0.001). An AI framework integrating DeepOxyMap and a Residual Generative Adversarial Network generated synthetic T1 maps directly from contrast-free oxygenation-sensitive CMR that demonstrated excellent agreement with native T1 values (r = 0.976, p < 0.001).