Deep Learning MR Image Denoising with Synthetic Phase and Noise

Motivation: Deep learning (DL) based complex image denoising (CpxDN) has shown advantages in noise suppression, but the phase images are often not stored in clinical settings. Goal(s): Generate synthetic phase image along with the magnitude image for CpxDN, and investigate its denoising performance. Approach: Two DL models were trained on 3027 paired images: one using magnitude pairs, and the other using complex-valued pairs with synthetic phases. Their denoising performances were evaluated quantitatively on another 63 paired test set, and qualitatively on 5 high noise test cases. Results: CpxDN with synthetic phase showed better denoising in the high level and structured noise scenarios. Impact: Synthetic phase allows convenient clinical deployment of complex DL denoising (CpxDN) models that shows advantage in high level and structured noise suppression. More clinical evaluation and optimization on CpxDN performance worth further investigation.