An Improved Deep Persistent Memory Network for Rician Noise Reduction in MR Images. (August 2022)

• Record Type: Journal Article
• Authors: Augustin, Anate Mary; Kesavadas, Chandrasekharan; Sudeep, P.V.
• Is Part Of: Biomedical signal processing and control. Volume 77(2022)
• Publication Date: 2022-08
• Subjects: Deep learning -- Denoising -- MemNet -- MRI -- Parallel imaging -- Rician -- Quantum ReLU; 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.2022.103736
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: An improved deep persistent meomory network is proposed for enhancing MRI images. The proposed network introduces nonlinearity via quantum-ReLU activation function. The proposed model utilizes a combined loss function with MMSE and SSIM losses. Experiments validate that the proposed method outperforms SoA methods. Our model has a better performance on stationary and non-stationary Rician distributed MR images. Abstract: Magnetic Resonance Imaging (MRI) is extensively employed in medical, scientific and investigative contexts today. Noise on the other hand, restricts the diagnostic utility of MR images by deteriorating their quality during acquisition. The noise in single coil magnitude MRI has stationary Rician distribution and images reconstructed with parallel MR-imaging techniques have non-stationary noise levels. Recently, deep learning models are finding ubiquitous employment in image restoration tasks, owing to their powerful capabilities in learning and solving inverse-problems. Nonetheless, only a few such techniques have been reported to suppress noise in MRI. In this paper, we propose a robust MR image denoising approach based on the concept of memory persistence. Accordingly, we improvised and optimized the deep model of memory networks by introducing a data sensitive activation function and a robust cost function, resulting in a compact design with improved noise filtering, feature preservation and enhanced performance. Experiments on real and synthetic data reveal that the proposed method outperformed state-of-the-art (SoTA) methods.