Noise dependent training for deep parallel ensemble denoising in magnetic resonance images. (April 2021)

• Record Type: Journal Article
• Authors: Aetesam, Hazique; Maji, Suman Kumar
• Is Part Of: Biomedical signal processing and control. Volume 66(2021)
• Publication Date: 2021-04
• Subjects: Magnetic resonance imaging -- Denoising -- Gaussian-impulse noise -- Deep learning -- Loss functions -- Laplacian distribution; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2020.102405
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A deep learning technique for Magnetic Resonance Image (MRI) denoising. Takes into account mixed Gaussian-impulse nature of noise. Derives a loss function for training from the Bayesian likelihood of Gaussian-Laplace distribution. Uses fully convolutional neural network with (CNN) based 3D residual learning strategy. Abstract: In this paper, a deep learning technique is proposed for the removal of Gaussian-impulse noise from Magnetic Resonance images (MRI). The proposed technique is inspired from the Bayesian maximum a posteriori (MAP) derivation of the Gaussian-impulse likelihood. A discriminative learning strategy under fully convolutional neural network (CNN) is used which focuses on the importance of loss layer during training. Residual learning is combined with 3D convolution for multi-dimensional extraction of image features from noisy data, on a wide range of noise levels. The problem of vanishing gradient in a very deep network is handled through the usage of a wide network, which is built by incorporating two parallel models (thereby resulting in decreased depth of the network). The approach is called ensemble because features are obtained along the two parallel paths working independently, using normal and dilated convolutions. Results on model convergence support advantages observed by these considerations. Experiments are conducted on synthetically corrupted MRI data and real spin-echo MRI sequence. Better visual and metric results as well as fast testing performance support the argument of boosted denoising capability against a majority of the benchmarks for MRI noise removal.