Development of Synthetic Parametric T2 Maps Using OS-CMR and Convolutional GANs

A dataset of 958 patients, paired of T2-weighted and OS-CMR images were used for training and testing. Various GANs were compared and the most performant was selected based on PSNR (Peak-Signal-to-Noise Ratio), SSIM (Structural-Similarity-Index), PCC (Pearson-Correlation-Coefficient), and MAE (Mean-Absolute-Error) metrics. The presence of signal intensity patterns within the generated image was then validated against ground truth using certified software and a custom-created Python code, with colour mapping illustrating the differences in signal intensity within the images.  T2 curves were then generated from the synthetic image and matched with a model-based T2 library in order to assess the correlation between the real and synthetic curves, ensuring the accuracy of the synthesized data. Using the model, multiple echo times were then merged to synthesize one image and decay curves were derived to generate a T2 map. Lastly, synthetic T2 decay curves derived from various normal and abnormal regions were compared to the original curves using correlation, MAE, RMSE, and R-squared.



Results:

Figure-1 illustrates our pipeline, starting with dataset preprocessing in. the R-GAN was selected for its superior performance in controlling vanishing gradients. The model achieved the highest PSNR (23.05 dB) and SSIM (0.4765), outperforming other models (Table-1). The synthesized images preserved critical diagnostic information using visual signal intensity analysis. Library-matched T2 curves showed a strong correlation with the synthetic curves (0.93) and an MAE of 0.56. The high correlation coefficient (0.97) and R-squared value (0.95) indicated that the synthesized data closely mirrored the original T2 data. T2 decay curves for edema ROIs closely matched the originals, with a low MAE of 8.97, a correlation of 0.99, and an RMSE of 8.97. An R-squared value of 0.85 confirmed the reliability of the synthesized data in replicating true edema signal behavior.



Conclusion: RGAN have the potential to synthesize a parametric T2 map from a single slice of OS-CMR. Further studies are needed to validate these findings.

Figure 1) Workflow for synthesizing edema maps from OS-CMR images Step 1: Dataset Preparation and Annotation - Matching slices of T2-weighted and OS-CMR images are identified. Preprocessing steps, including normalization, scaling, augmentation, and applying window techniques. Step 2: GAN Training - Various GAN models are trained to find the best match for generating synthetic T2-weighted images from input data. The proposed GAN model with residual blocks, referred to as RGAN, demonstrates the best results in capturing the nuances of edema. Step 3: Validation Metrics - The synthesized images are evaluated against ground truth using various metrics like PSNR , SSIM and PCC to ensure their accuracy and fidelity. Step 4: Signal Intensity Verification - The signal intensity of the synthesized images is compared with reference images using a Python script and certified CMR analysis software (cvi42) to test for the preservation of critical diagnostic information. Step 5: T2 Curve and Map Generation - T2 curves are predicted using a model based on MRI parameters, and we assess the correlation between the real and synthetic curves, ensuring the accuracy of the synthesized data. Step 6: Edema Case Testing and Analysis - The final synthesized images are tested on a case with known edema. T2 decay curves and maps are generated by replacing the synthesized image within a folder containing other T2-weighted image slices. These synthetic images are compared with the original images across various echo times (TEs) to validate their reliability. The resulting T2 decay curves for edema and non-edema regions are compared with original data to calculate RMSE, MAE, and correlation, confirming the synthetic images' ability to replicate the diagnostic features of edema in CMR.
new fig GAN.pdf