A 3D Wide Residual Network with Perceptual Loss for Brain MRI Image Denoising

Image denoising is an essential pre-processing step in medical image analysis. It brings about an improvement in the accuracy of disease diagnosis (done by analysis of medical images), and helps to determine the correct course of treatment plan. Over the past few years, Convolutional Neural Networks (CNNs) have been proven to be extremely effective in image processing and computer vision applications. In this work we explore efficient wide residual CNNs for denoising human brain magnetic resonance images. Many state-of-the-art CNN based image denoising methods use squared Euclidean distance for training the neural network. This causes deep networks to produce over-smoothed image outputs, losing structural or anatomical details. Also, the details of structures (found in images) are often lost in the deep hierarchy of convolution layers. Unlike such existing methods, in this paper, we use perceptual loss alongside squared Euclidean distance for training our network. This helps restoring images that are visually desirable, in the sense that they restore considerably more anatomically refined features, and are not just similar to the ground-truth pixel intensity. Our experiments show that the proposed convolutional network surpasses the state-of-the-art methods of rician MRI denoising, and obtains better quality denoised brain MR images as compared to the state-of-the-art.