Figure 1.
Illustration of NLM filtering effect on small high-contrast particles.
(a) Synthesized checkerboard image with one-pixel particle details; (b) Gaussian noise-corrupted image; (c) NLM-filtered image; (d) NLM-calculated weights of pixels in the search window as enclosed by the square in (b).
Table 1.
Optimal h values for different denoising algorithms, image types, radius of patches (rp), and a search window with a radius (rs) of 5.
Figure 2.
PSNR comparison of RNLM and RNLM-CPP algorithms under varying noise levels (ranging from 1% to 9% with an increase of 2%) for different image types (T1w, T2w, and PD) and patch sizes (radius of 1, 2, and 3).
Figure 3.
Comparison of RNLM and RNLM-CPP algorithms on denoising simulated T1w images.
Top row, from left to right: noisy image with 5% of Rician noise, denoised results with different algorithms. Second row, from left to right: zoomed part of the corresponding images in the top row, the dotted boxes indicate the local areas around manually-defined particles. Bottom row, from left to right: T1w noise-free image and corresponding image residuals.
Figure 4.
Comparison of RNLM and RNLM-CPP algorithms on denoising simulated T2w images.
Top row, from left to right: noisy image with 5% of Rician noise, denoised results with different algorithms. Second row, from left to right: zoomed part of the corresponding images in the top row, the dotted boxes indicate the local areas around manually-defined particles. Bottom row, from left to right: T2w noise-free image and corresponding image residuals.
Figure 5.
Comparison of RNLM and RNLM-CPP algorithms on denoising simulated PDw images.
Top row, from left to right: noisy image with 5% of Rician noise, denoised results with different algorithms. Second row, from left to right: zoomed part of the corresponding images in the top row, the dotted boxes indicate the local areas around manually-defined particles. Bottom row, from left to right: PDw noise-free image and corresponding image residuals.
Table 2.
LPSNR results for quantitative comparison of RNLM and RNLM-CPP algorithms with parameters (rp = 1, rs = 5, = 4, and
= 5) for T1w, T2w and PDw images.
Table 3.
LSSIM results for quantitative comparison of RNLM and RNLM-CPP algorithms with parameters (rp = 1, rs = 5, = 4, and
= 5) for T1w, T2w and PDw images.
Figure 6.
Comparison of RNLM and RNLM-CPP algorithms on denoising real T1 image.
Top row, from left to right: real T1w image and denoised results from different algorithms. Second row, from left to right: zoomed image in the square of the top-left image. Third row: intensity profile of noisy image and denoised images located on the cyan line across the particle. Bottom row, from left to right: corresponding residuals.
Figure 7.
Comparison of RNLM and RNLM-CPP algorithms on denoising real diffusion-weighted image.
Top row, from left to right: real diffusion-weighted image and denoised results from different algorithms. Second row, from left to right: zoomed image in the square of the top-left image. Third row: intensity profile of noisy image and denoised images located on the cyan line across the particle. Bottom row, from left to right: corresponding residuals.