Deep Neural Networks for Noise Reduction and Bias Removal in MR Diffusion Signal

Abstract

The purpose of this study is to assess the potential of deep neural networks, trained by unsupervised learning, for diffusion weighted imaging (DWI) data modeling and denoising. DWI data were modeled by a biexponential model and Rician bias was corrected for. Deep neural networks that estimate the magnetic resonance (MR) diffusion-weighted signal decay were trained on simulated signal data. Results for simulated data with known σg and estimated σg were compared, where known σg was the most suitable method. Furthermore, a deep neural network trained directly on patient prostate data was used to denoise images. The method of using deep neural networks was compared with OBSIDIAN, which is a model-based, iterative fitting procedure. The deep neural network showed an improvement of image quality with respect to the raw data, but did not have the same quality as OBSIDIAN. Using the trained deep neural network on the same patient data resulted in a runtime of 1.9 ms. The results showed that there is some potential in using deep neural networks for DWI data modelling and denoising, but further optimization is needed.