Compressed Sensing MRI Reconstruction Using Convolutional Dictionary Learning and Laplacian Prior

Abstract

In compressed sensing magnetic resonance imaging (CS-MRI), applications of dictionary learning techniques have craved a decade long way with the development of methods like K-SVD, matching pursuits, etc. Dictionary learning methods are particularly useful in context of input signal adaptability. The data acquisition process of MRI is noisy in nature with various types of noise associated, like Rician, Gaussian, Rayleigh noise, motion artefacts like breathing artefacts, etc. In this context, training a dictionary directly with the noisy training samples may lead to an inefficient dictionary. Moreover, complexity and size of the constructed dictionary may be very big. This paper proposes a Laplacian sparse dictionary (LSD) technique for obtaining a concise and more representative dictionary which utilizes the concepts of manifold learning and double sparsity for MR image. This can be utilized to reconstruct an MR image using any of the existing compressed sensing methodology. The method along with online convolutional dictionary learning (CDL) has been demonstrated in this manuscript. Keeping in mind increased efficiency and reduced reconstruction time, the proposed method attempts to tackle the problem of MR image reconstruction. The results obtained from the proposed method have been compared with traditional CS-MRI methods using metrics—PSNR and SSIM.