Generating R2^* maps from T1W and T2W images using image-to-image translation for Parkinson's disease

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Quantitative R2* mapping reveals iron deposition and tissue changes, aiding Parkinson's disease (PD) management.
• Clinical use of R2* maps is limited by time constraints and susceptibility artifacts.

Purpose of the Study:

• Evaluate the feasibility of using generative adversarial networks (GANs) to synthesize R2* maps.
• Generate R2* maps from readily available T1-weighted (T1W) and T2-weighted (T2W) MRI images.

Main Methods:

• Developed a GAN model to synthesize R2* maps from T1W and T2W images.
• Compared GAN performance against a 2D Unet model using metrics like NMSE, PSNR, SSIM, FSIM, and RMSE.
• Assessed diagnostic efficacy using Area Under the ROC Curve (AUC) for distinguishing PD from healthy controls, focusing on the substantia nigra pars compacta (SNpc).

Main Results:

• The GAN model outperformed the 2D Unet model.
• High correlations (0.75-0.87) observed between synthetic and real R2* maps in internal datasets.
• AUC values of 0.79 (synthetic) and 0.80 (real) for internal data; 0.84 for synthetic data in external datasets.
• Longitudinal analysis showed significant correlation between ∆R2* and ∆UPDRS in SNpc and SNpr for the PD group.

Conclusions:

• Synthetic R2* maps correlate well with real maps and show promise for PD diagnosis and assessment.
• GAN-based R2* map synthesis offers a potential solution to overcome limitations of traditional R2* mapping.