Weighted Low-Rank and Sparse Matrix Decomposition Models for Separating Background and Foreground in Dynamic MRI

Abstract

AbstractDynamic Magnetic Resonance Imaging (MRI) allows non-invasiveness, non-ionization monitoring, and study of the functional and anatomical changes of internal body structure over time. For the improved disease diagnosis, Machine Learning (ML) techniques are widely employed for the background and foreground separation in the dynamic MRI. Separation using Neural Networks has several limitations like high processing time and the requirement of a large number of datasets. Recently the Low-Rank and Sparse Decomposition (LRSD) has received a lot of attention in the foreground–background separation in dynamic MRI. Traditional LRSD approaches, on the other hand, have the issues such as estimating exact low-rank structure and sensitivity to the noise. To deal with these problems, this paper proposes the Low-Rank and Sparse Decomposition models such as the Weighted Nuclear Norm Minimization model (WNNM) and Weighted Schatten p-Norm Minimization model (WSNM) for the foreground–background separation in the dynamic MRI to approximate the low-rank function. Extensive experiments have been conducted using the above models on dynamic MRI data set and performance is compared with Robust PCA (RPCA). The results show, both intuitively and numerically, that WSNM outperforms RPCA and WNNM in separating the background and the foreground information in dynamic MRI, with greater Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) values.KeywordsMagnetic resonance imaging (MRI)Machine learning (ML)Low-rank and sparse decomposition (LRSD)Robust principal component analysis (RPCA)Weighted nuclear norm minimization (WNNM)Weighted schatten p-norm minimization (WSNM)Peak signal-to-noise ratio (PSNR)Structural similarity index (SSIM)