Unsupervised deep learning for denoising diffusion-weighted images with noise-correction loss functions

Motivation: Since the noisy magnitude MR data generally follows Rician distribution, using the noisy images and network’s output to construct unsupervised learning’s loss function for denoising will lead to a biased estimation, especially for DW images which suffers from the lower SNR. Goal(s): To address the noise bias issue. Approach: We proposed two noise-correction loss functions for unsupervised denoising of DW images, based on DIP and the characteristics of Rician distribution. Results: The experimental results on simulated and in-vivo data demonstrated that the proposed loss functions effectively corrected the signal-dependent noise bias and improved the accuracy of unsupervised learning-based DW images denoising method. Impact: Firstly, we proposed two noise-correction loss functions and validate their effectiveness in denoising DW images. Secondly, the proposed loss functions are not limited to DW images and can be directly applied to other modality MR images.