A future cmr workflow integrating DeepOxyMap and residual GAN–based synthetic T1 and T2 mapping to differentiate ischemic, non-ischemic, amyloid, and edema patterns
F Lotfikazemi, M B Benovoy, M C Chetrit, L H Haririsanati, J R Rafiee, M F FriedrichAbstract
Introduction
CMR remains limited by multiple acquisition sequences and contrast injection [1][2]. Needle-free Oxygenation-Sensitive CMR (OS-CMR) potential for contrast-free assessment of myocardial tissue[3]. To address this limitation, We propose a novel unified AI framework combining DeepOxyMap and a Residual GAN to extract diagnostic patterns and generate quantitative T1/T2 maps from OS-CMR. We hypothesized that this contrast-free approach could transform CMR workflows.
Methods
OS-CMR images from two prospective studies were used to develop DeepOxyMap for diagnostic heatmap generation (LGE ground truth), and a novel R-GAN to synthesize quantitative T1/T2 maps directly from OS-CMR. Clinical validation included:(1) Signal intensity comparison between real and synthetic; (2) EPG simulation based on MRI-dictionary matching for T1/T2 curve analysis validated by quantitative agreement metrics ;(4) Curve and map assessment in CVI42 software.
Results
DeepOxyMap developed by 197 OS-CMR on different clinical cases to extracted diagnosis patterns with achieving AUC= 0.94 and Dice similarity > 0.89 compared to ground-truth. After patterns extracted, our novel R-GAN apply on 2044 and 2089 images to synthesis parametric maps from OS-CMR, achieving SSIM 0.672–0.739, PSNR 21.56–24.95 dB, PCC 0.862–0.923 on T2 images and SSIM 0.810, PSNR 16.01 dB, PCC 0.748 on T1 images. Myocardial signal-intensity similarity reached 97% in amyloid, 95% in ischemic/non-ischemic, and 98% in edema cases. EPG-based T1 simulations showed strong agreement between synthetic and real images across all groups. In Amyloid: MAE=0.0037, MSE=0.0010, RMSE=0.0068, R²=0.9966, Pearson-correlation = 0.9999. In ICMP: MAE = 0.0054, MSE = 0.0002, RMSE = 0.0108, R² = 0.9946, Pearson-correlation≥0.9998. In NICMP: MAE = 0.0094, MSE = 0.0004, RMSE = 0.0185, R²=0.9681, Pearson-correlation≥0.9993. EPG-based T2-curve in edema: MAE=0.0007, RMSE =0.0008, R² values≥0.87, Pearson-correlations ≥0.99. T1 and T2 curves on different ROIs between both real and synthesis images shows promising results, In Amyloid group: MAE=0.0041,MSE=0.0001, RMSE=0.0078, R²=0.9979, Pearson-correlation=0.9998, Spearman-correlation=1.0000. In ICMP group:MAE=0.0062, MSE=0.0002, RMSE= 0.0114, R²=0.9941,Pearson-correlation= 0.9997, Spearman-correlation=1.0000.In NICMP group: MAE=0.0088,MSE=0.0004,RMSE= 0.0176, R²= 0.9724, Pearson-correlation = 0.9994, Spearman-correlation=1.0000. T2-curves in edema-ROIs: correlation ≥0.99, MAE=0.0004, RMSE=0.004, R² ≥0.99. The synthetic T1/T2 colour-maps were compared with the real maps using AHA segmentation.
Conclusion
This study introduces the first contrast-free AI framework capable of extracting pathology-specific patterns and synthesizing quantitative T1/T2 maps directly from a single acquisition to challenges contrast-dependent CMR-workflows and supports rapid, non-invasive myocardial tissue characterization.