Deep variational magnetic resonance image denoising via network conditioning

In this paper, we propose a variational approach towards denoising magnetic resonance images (MRI) corrupted by spatially variant and signal-dependent Rician noise in a deep learning framework. To obtain a mathematically sound inference network, approximate variational posteriors are designed keeping in mind the Rician nature of noise. The proposed work tackles the denoising problem in several different ways. Firstly, the prior assumption on data in the variational posterior is motivated by the heavy-tailed marginal distribution of image gradients in natural images. This is captured by the sparsity promoting hyper-Laplacian prior on MR data. Similarly, median absolute deviation under Gaussian prior helps in the estimation of noise in the variational lower bound of marginal log likelihood term. Secondly, noise estimation from the background regions of the noisy data under the assumption of Rayleigh distribution prevents the addition of extra sub-network for the estimation of spatially variant noise level parameters. Thirdly, feature-wise transformation of intermediate layers is performed using anatomical planes segmentation maps (APSM) for context-based network conditioning. Here, affine transformation parameters generated from APSM are modulated with the input features for spatial feature transformation. Fourthly, to capture the long-range dependencies lost in deeper convolutional layers, multiscale global feature fusion block (GFFuB) is used. Lastly, experimental results over synthetically corrupted MR data and real data obtained from MR scanners suggest the potential utility of the proposed model in real time. [Display omitted] •A Bayesian-motivated variational autoencoder is used to understand the image generation process.•Variational lower bound estimates both the noise level parameter and the true signal amplitude.•Instead of a trainable model, maximum likelihood estimator (MLE) is used to obtain an unbiased noise level parameter.•Context-based network conditioning is facilitated by spatial transformation of input features obtained from anatomical planes segmentation maps.