Denoising of 3D magnetic resonance images with multi-channel residual learning of convolutional neural network

摘要

The denoising of magnetic resonance (MR) images is a task of great importancefor improving the acquired image quality. Many methods have been proposed inthe literature to retrieve noise free images with good performances. Howerever,the state-of-the-art denoising methods, all needs a time-consuming optimizationprocesses and their performance strongly depend on the estimated noise levelparameter. Within this manuscript we propose the idea of denoising MRI Riciannoise using a convolutional neural network. The advantage of the proposedmethodology is that the learning based model can be directly used in thedenosing process without optimization and even without the noise levelparameter. Specifically, a ten convolutional layers neural network combinedwith residual learning and multi-channel strategy was proposed. Two trainingways: training on a specific noise level and training on a general level wereconducted to demonstrate the capability of our methods. Experimental resultsover synthetic and real 3D MR data demonstrate our proposed network can achievesuperior performance compared with other methods in term of both of the peaksignal to noise ratio and the global of structure similarity index. Withoutnoise level parameter, our general noise-applicable model is also better thanthe other compared methods in two datasets. Furthermore, our training modelshows good general applicability.