MRI Denoising Using Pixel-Wise Threshold Selection

Abstract

Magnetic resonance imaging (MRI) has emerged as a promising technique for non-invasive medical imaging. The primary challenge in MRI is the trade-off between image visual quality and acquisition time. Current MRI image denoising algorithms employ global thresholding to denoise the whole image, which leads to inadequate denoising or image distortion. This study introduces a novel pixel-wise (localized) thresholding approach of singular vectors, obtained from singular value decomposition, to denoise magnetic resonance (MR) images. The pixel-wise thresholding of singular vectors is performed using separate singular values as thresholds at each pixel, which is advantageous given the spatial noise variation throughout the image. The method presented is validated on MR images of a standard phantom approved by the magnetic resonance accreditation program (MRAP). The denoised images display superior visual quality and recover minute structural information otherwise suppressed in the noisy image. The increase in peak-signal-to-noise-ratio (PSNR) and contrast-to-noise-ratio (CNR) values of ≥ 18% and ≥ 200% of the denoised images, respectively, imply efficient noise removal and visual quality enhancement. The structural similarity index (SSIM) of ≥ 0.95 for denoised images indicates that the crucial structural information is recovered through the presented method. A comparison with the standard filtering methods widely used for MRI denoising establishes the superior performance of the presented method. The presented pixel-wise denoising technique reduces the scan time by 2-3 times and has the potential to be integrated into any MRI system to obtain faster and better quality images.

Keywords: Magnetic resonance imaging; SF-SVD; contrast-to-noise ratio; denoising; image denoising; peak-signal-to-noise ratio; pixel-wise noise threshold selection; pixel-wise thresholding; singular value decomposition; structural similarity index.

FIGURE 1.

Flowchart of current denoising methods vs presented pixel-wise denoising method. AFHB: Adaptive fuzzy hexagonal bilateral, PSNLM: Pre-smooth non-local means.

FIGURE 2.

Flowchart of the presented pixel-wise thresholding (PWT) method.

FIGURE 3.

Plot of $S$ vs $l$ at different pixels for images of (a) Slice-1, (b) Slice-2, (c) Slice-3, and (d) Slice-4. $S$ at different pixels saturate at different singular values depending on the PSNR of that pixel. The vertical lines belong to the threshold singular value.

FIGURE 4.

Denoising results of MRI scan of the ACR phantom for slices 2 (first row), 8 (second row), and 10 (third row) respectively. The noisy and reference images are obtained after averaging 3 and 16 identical scans respectively. The denoised images display improved image quality and enhanced visibility.

FIGURE 5.

Denoising results of MRI scans of ACR Phantom for slices 3 and 13 at scan numbers 5,7, and 9, respectively. The results are consistent across different scans.

FIGURE 6.

Denoising results of MRI scan of the ACR phantom for slices 1,7 and 16 respectively. The highlighted area in the figure illustrates the recovered features after denoising.

FIGURE 7.

Comparison of PWT with the adaptive fuzzy hexagonal bilateral and PSNLM filtering of slices 1 (first row), 7 (second row), and 16 (third row). Adaptive fuzzy hexagonal bilateral filtering is able to reduce noise but fails to recover features highlighted in the red ellipse. PSNLM filtering is able to recover features but they are difficult to identify without knowing they existed before. PWT is able to reduce noise as well as recover features. AFHB: Adaptive fuzzy hexagonal bilateral, PSNLM: Pre-smooth non-local means.

FIGURE 8.

Comparison of PWT with the bilateral and NLM filtering of slices 1 (first row), 7 (second row), and 16 (third row). Bilateral filtering is able to reduce noise but fails to recover features highlighted in the red ellipse. NLM filtering is able to recover features but they are difficult to identify without knowing they existed before. PWT is able to reduce noise and recover features.