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

Previous studies show that quantitative R2^* mapping can reveal iron deposition and tissue alterations, potentially aiding Parkinson's disease (PD) management. However, R2^* maps are not commonly used in clinical practice due to the extra time required and sensitivity to susceptibility artifacts.

In this work, we propose to evaluate the feasibility of using generative-adversarial-networks (GANs) for synthesis of R2^* maps from T1-weighted (T1W) and T2-weighted (T2W) images.

A GAN model was developed to synthesize R2^* maps from T1W and T2W images. 572 internal participants and 268 external participants were included. The internal-dataset was divided into training (344), validation (114), and test data (114), while the external-dataset was an independent test-set. The performance of the proposed model was compared with a 2D Unet model without adversarial loss. The performance of the two models was evaluated using normalized mean square error, peak signal to noise ratio, the structure similarity index measure (SSIM), feature similarity index (FSIM), root mean square difference, average absolute difference, and relative error. Pearson method was used to assess the correlation between synthetic and real values. The area-under-the-receiver-operating-characteristic-curve (AUC) was calculated to evaluate the diagnostic efficacy of R2^* maps in distinguishing PD from healthy controls (HC), with a focus on the substantia nigra pars compacta (SNpc).

The proposed model performed better than the 2D Unet model. In internal test-set, high correlations were observed between synthetic and real R2^* maps across various brain regions, with coefficients ranging from 0.75 to 0.87. The AUC was 0.79 and 0.80 for synthetic and real maps (p = 0.76), respectively. In external test-set, the AUC was 0.84 for synthetic R2^* maps. Longitudinal analysis showed a positive correlation between ∆R2^* and ∆UPDRS (Unified-Parkinson's-Disease-Rating-Scale) in SNpc (R = 0.69, p = 0.01) and substantia nigra pars reticulata (SNpr) (R = 0.64, p = 0.02) for PD group.

The synthetic R2^* maps demonstrated good correlation with real maps and performed well in diagnosing and evaluating PD in both internal and external datasets, indicating their potential value for PD diagnosis and assessment.