R-GAN: Contrast-Free Generation of Native T1 Parametric Maps from OS-CMR for Myocardial Ischemic and Non-Ischemic Detection

Presenting Author(s)

• 

  Faezeh Llotfikazemi

  PhD candidate
  McGill University
  montreal, Quebec, Canada

Primary Author(s)

• 

  Faezeh Llotfikazemi

  PhD candidate
  McGill University
  montreal, Quebec, Canada

Co-Author(s)

• 

  Matthias G. Friedrich, MD, FSCMR

  Professor, Chief, Cardiovascular Imaging, McGill University Health Centre
  McGill University Health Center, Research Institute of the McGill University Health Centre
  Montreal, Quebec, Canada
• MB

  Mitchel Benovoy, PhD

  PhD
  Area 19 Medical Inc.
  Montreal, Quebec, Canada
• LH

  Leila Haririsanati, MD, MSc

  Researcher
  McGill University Health Center
  Montreal, Quebec, Canada
• MR

  Moezedin javad Rafiee, MD

  Research associate
  Research Institute of the McGill University Health Center
  Montreal, Quebec, Canada
• MC

  Michael Chetrit, MD

  Assistant professor
  McGill University Health Center
  Montreal, Quebec, Canada

Background: Myocardial fibrosis is a hallmark of both ischemic and non-ischemic cardiomyopathies and is typically assessed using T1 mapping or late gadolinium enhancement (LGE). However, LGE requires intravenous contrast, which carries safety concerns, particularly in patients with renal impairment, and limits repeat follow-up. We present a Residual Generative Adversarial Network (R-GAN) that generates native T1 parametric maps from contrast-free oxygenation-sensitive CMR (OS-CMR). This approach enables non-invasive detection of myocardial fibrosis, supporting differentiation between ischemic and non-ischemic cardiomyopathies.

Methods: A total of 2,044 paired OS-CMR and native T1-W images were used to train and evaluate the R-GAN model. Preprocessing included rigid registration, windowing, normalization, and resizing to 256×256 pixels. Data augmentation involved rotations (±20° to ±180°) and flips, creating aligned OS–T1 pairs. Model performance was compared with Pix2Pix with or without augmentation using SSIM, PSNR, and Pearson correlation coefficient (PCC). Clinical validation was performed in representative ischemic and non-ischemic cases. Signal intensities (SI) were measured from ROIs in fibrotic and remote myocardium. Synthetic and reference T1 recovery curves were extracted using CVI42 and custom Python scripts and validated against Extended Phase Graph (EPG) simulations. Synthetic and reference T1 colour maps were visually compared in CVI42.

Results: R-GAN outperformed Pix2Pix, generating higher-fidelity T1 maps (SSIM = 0.805, PSNR = 16.1 dB, PCC = 0.748). ROI analysis confirmed preserved tissue contrast, with SI differences in fibrotic regions of ~7–8% compared to reference images. T1 recovery curves from synthetic and reference maps demonstrated strong agreement:

• 
  Ischemic cardiomyopathy: MAE = 0.019, MSE = 0.0005, RMSE = 0.023, R² = 0.99, Pearson = 0.999, Spearman = 1.000.
  
  


• 
  Non-ischemic cardiomyopathy (Case 1): MAE = 0.029, MSE = 0.0023, RMSE = 0.048, R² = 0.97, Pearson = 0.986, Spearman = 0.929.
  
  


• 
  Non-ischemic cardiomyopathy (Case 2): MAE = 0.033, MSE = 0.0085, RMSE = 0.092, R² = 0.87, Pearson = 0.941, Spearman = 0.976. (Table 1)
  
  

Visual inspection in CVI42 confirmed high concordance in curve evolution and plateau across all cases. Additional validation using Extended Phase Graph (EPG) simulations further confirmed synthetic curve fidelity (MAE = 0.007, MSE = 0.0002, RMSE = 0.013, R² = 0.996, Pearson = 0.998, Spearman = 1.000). Synthetic T1 color maps closely matched reference maps, accurately depicting regional fibrosis patterns (Figure 1).

Conclusion: R-GAN generates high-fidelity native T1-maps from OS-CMR without the need for contrast administration, outperforming pix2pix GANs and demonstrating robust fidelity in ischemic and non-ischemic cardiomyopathy. This contrast-free, repeatable approach enables non-invasive myocardial fibrosis imaging and holds promise for improving accessibility and follow-up in patients where contrast-enhanced scans are not feasible.

Table 1. Quantitative Similarity Metrics Between Original and Synthesized T1 Curves Acrossischemic and non-ischemic Cardiac Pathologies

Pathology Type	Evaluation Method	MAE	MSE	RMSE	R²	Pearson	Spearman
Ischemic	Python Script	0.0191	0.0005	0.0228	0.9943	0.9993	1.0000
	CVI42 (Visual)	Excellent alignment in curve shape and amplitude across 2 ROIs
Non-Ischemic (Case1, HCM)	Python Script	0.0294	0.0023	0.0483	0.9720	0.9863	0.9286
	CVI42 (Visual)	High visual agreement in curve evolution and plateau
Non-ischemic (Case 2, DCM)	Python Script	0.0325	0.0085	0.0920	0.8668	0.9406	0.9762
	CVI42 (Visual)	High visual agreement in curve evolution and plateau
Non-Ischemic- EPG Validation	EPG, MRI Dictionary Based Simulation	0.0069	0.0002	0.0125	0.9958	0.998	1.000

Figure 1. Workflow and validation of R-GAN synthetic native T1 mapping from OS-CMR. Representative outputs, ROI-based validation, and recovery curves show strong agreement between synthetic and reference T1 values. Synthetic T1 color maps preserved spatial distribution of fibrosis in ischemic and non-ischemic cardiomyopathy.