Supervised denoising of diffusion-weighted magnetic resonance images using a convolutional neural network and transfer learning

Abstract

In this paper, we propose a method for reducing thermal noise in diffusion-weighted magnetic resonance images (DWI MRI) of the brain using a convolutional neural network (CNN) trained on realistic, synthetic MR data. Two reference methods are considered: a) averaging of repeated scans, a widespread method used in clinics to improve signal-to-noise ratio of MR images and b) the blockwise Non-Local Means (NLM) filter, one of the post-processing methods frequently used in DWI denoising. To obtain training data for transfer learning, the effects of echo-planar imaging (EPI) – Nyquist ghosting and ramp sampling – are modelled in a data-driven fashion. These effects are introduced to the digital phantom of brain anatomy (BrainWeb). Real noise maps are obtained from the MRI scanner with a brain-DWI-designed protocol and later combined with simulated, noise-free EPI images. The Point Spread Function is measured in a DW image of an AJR-approved geometrical phantom. Inter-scan patient movement is captured from a brain scan of a healthy volunteer using image registration. The denoising methods are applied to the simulated EPI brain images and in real EPI DWI of the brain. The quality of denoised images is evaluated at several signal-to-noise ratios. The characteristics of noise residuals are studied thoroughly. A diffusion phantom is used to investigate the influence of denoising on ADC measurements. The method is also evaluated on a GRAPPA dataset. We show that our method outperforms NLM and image averaging and allows for a significant reduction in scan time by lowering the number of repeated scans. We also analyse the trained CNN denoisers and point out the challenges accompanying this denoising method.