Denoising of Rician corrupted 3D magnetic resonance images using tensor-SVD. (July 2018)

• Record Type: Journal Article
• Authors: Khaleel, Hawazin S.; Mohd Sagheer, Sameera V.; Baburaj, M.; George, Sudhish N.
• Is Part Of: Biomedical signal processing and control. Volume 44(2018)
• Page Start: 82
• Page End: 95
• Publication Date: 2018-07
• Subjects: MR imaging -- Rician noise -- Tensor-SVD -- Low rank recovery; 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.2018.04.004
• Languages: English
• ISSNs: 1746-8094

Abstract

Graphical abstract: Highlights: A new method for denosing MR images is formulated for removing the conventional Rician noise in the t -SVD framework. The low rankness of three dimensional MR images was exploited by grouping similar cubic patches as tensors. The performance of the algorithm on synthetic and real MR images were tested using different metrics. The proposed method was found to be more effective for tumour detection of noisy MR images than the counterparts. Abstract: In this paper, we propose a new method for denoising the volumetric magnetic resonance (MR) images degraded with Rician noise. Taking into account the multi-frame (multi-linear) nature, the proposed method formulates an unsophisticated approach by contemplating the MR data as third order tensor. Since the Rician noise is signal dependent, variance stabilization technique (VST) is applied to transform it as additive noise. The cubic patches extracted from 3D MR images are grouped as tensors which exhibit low-rank property. Thus, denoising problem is modelled as a low-rank tensor approximation of grouped tensors, solved by minimizing the tensor nuclear norm (TNN) in tensor -singular value decomposition ( t -SVD) framework. Each denoised tensors are weighted-averaged to obtain the final denoised data. The efficiency of proposed algorithm is compared with the state of art techniques and has exhibited substantial improvement in terms of quality metrics such as PSNR, SSIM and EPI for synthetic MR images. The algorithm performance is assessed for real MR images with no reference quality metric viz. sharpness index (SI) and have shown superior results. Moreover, the effectiveness of proposed algorithm for MR image segmentation is evaluated. As observed from results, the accuracy of segmentation with regards to kappa coefficient is improved by 1–5% after applying the proposed denoising algorithm.