A Conditional Generative Adversarial Network (cGAN) Based 2D MP-RAGE MR Image Synthesis Method

Purpose: A deep learning framework based on a conditional Generative Adversarial Network (cGAN) was developed to synthesize high-contrast Magnetization Prepared Rapid Gradient Echo (MP-RAGE) images from common spin-echo MR imaging sequences. This framework utilizes combinations of inputs from T1-weighted (T1-w), T2-weighted (T2-w), and Proton Density-weighted (PD-w) images. The primary objective was to augment the diversity of clinical data by capitalizing on the inherent advantages of MP-RAGE imaging, such as superior contrast, while mitigating its susceptibility to metal artifacts.Methods and Materials: A cGAN image synthesis model, incorporating a U-Net-based generator and a Patch GAN discriminator, was developed. The training was conducted across four distinct configurations, employing combinations of T1-w, T2-w, and PD-w images as inputs to synthesize MP-RAGE images, with and without Proton Density (PD) information, designated as PD-0 and PD-1, respectively. For training, data from 51 patients, comprising 8,160 slices, were used, following a training-to-validation ratio of 90:10. For prediction, data from 14 patients, comprising 2,240 slices, were utilized. The efficacy of the synthesized MP-RAGE images was evaluated using a suite of quantitative metrics, including Mean Absolute Error (MAE), Normalized Cross-Correlation (NCC), Percentage Mutual Information (PMI), and Structural Similarity Index (SSIM). Additionally, a Freesurfer brain segmentation task was performed on both synthesized and ground truth brain images, with the fidelity of synthesized images being indirectly assessed by the calculated Dice coefficient. Results: It was observed that the cGAN-synthesized MP-RAGE images exhibited comparable contrast to the ground truth in the axial view. A decrease in input channel numbers resulted in diminished contrast between certain anatomical structures in the synthetic MP-RAGE images, albeit within an acceptable range. The MAE approached (0.02±0.01), the PMI for two Three-in-One-out synthesis approached(0.76±0.07), the NCC was about (0.91±0.05), and the SSIM was about (0.9±0.1). The Freesurfer segmentation results showed desirable Dice coefficients (mostly above 0.8) for different kinds of inputs, except the One-in-One-out T1-w synthesis. Conclusion: The cGAN framework developed in this study has proven to be a robust and versatile tool for synthesizing high-contrast MP-RAGE images, even in scenarios with single-channel input images. The Freesurfer segmentation results demonstrated that the synthesized MP-RAGE images are highly similar to the ground truth in segmentation tasks, underscoring the potential clinical and research value of the proposed image synthesis model.