A no-reference metric to assess quality of denoising for Magnetic Resonance images. (September 2021)

• Record Type: Journal Article
• Authors: Simi, V.R.; Reddy Edla, Damodar; Joseph, Justin
• Is Part Of: Biomedical signal processing and control. Volume 70(2021)
• Publication Date: 2021-09
• Subjects: Denoising -- Image quality analysis -- No-reference image quality metrics -- Image smoothing; 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.2021.102962
• Languages: English
• ISSNs: 1746-8094

Abstract

Graphical abstract: Highlights: A no-reference metric for assessing quality of denoised MR images is proposed. The metric has good correlation with subjective quality scores. Computationally fast. Collectively reflects quality degradation due to residual noise and the blur at edges. Application in performance evaluation of denoising algorithms and parameter optimisation. Abstract: Performance evaluation of algorithms used for denoising real-time Magnetic Resonance (MR) images and the selection of their operational parameters are difficult as noise-free ground-truth is unavailable. No-reference metrics that can reflect the quality of denoised images with respect to the strength of residual noise and inadvertent blur at edges are required in this context. A no-reference metric, termed as Objective Measure of Quality of Denoised Images (OMQDI), for assessing the quality of denoised MR images is proposed in this paper. OMQDI is a sum of two quality factors termed as Edge-Preservation Factor (EPF) and Noise Suppression Factor (NSF). EPF is computed from the sharpness of edges in the noisy input and denoised images and NSF is computed from the noise power in them. The sharpness is computed in the Wavelet domain. The sharpness is the nonlinearly weighted sum of cumulative log-energy of wavelet coefficients from three different levels of decomposition. Cumulative log-energy at any particular level of decomposition is computed from log-energy from LH, HL and HH sub-bands corresponding to that level of decomposition. The noise power is estimated from the global mean of local grey level variance with the help of a non-parametric model. The Pearson's correlation with the subjective quality rating exhibited by No-Reference Structure Similarity (NRSS), Sparsity and Dominant orientation-based Quality Index (SDQI), MetricQ, Anisotropic Quality Index (AQI), Optimum Denoising Index (ODI) and OMQDI on 100 test datasets are 0.245 ± 0.1032, 0.5757 ± 0.2735, 0.5615 ± 0.2959, 0.6566 ± 0.2361, 0.8391 ± 0.085 and 0.9619 ± 0.0293, respectively. The computational time (in sec) of NRSS, SDQI, MetricQ, AQI, ODI and OMQDI are 0.0777 ± 0.0032, 0.1998 ± 0.1155, 2.4347 ± 0.0757, 5.4859 ± 1.4862, 0.2914 ± 0.0508 and 0.1382 ± 0.0164, respectively. OMQDI has good agreement with the subjective quality rating and it is fast in computation.