Chi-square risk estimation helps reduce noise in MR images

A combination of the established denoising algorithm NeighShrink and chi-square unbiased risk estimation (CURE) could reduce noise in magnetic resonance (MR) images more effectively than traditional methods, according to research published in Artificial Intelligence in Medicine.

While MRI is a critical tool for diagnosing and treating disease, it’s not without its faults, first author Chang-Jiang Zhang and colleagues at Zhejiang Normal University in Jinhua, China, said in the journal. The visual quality of an MRI can be easily degraded by stochastic variation during the acquisition procedure, rigid and nonrigid body movements and artifacts between tissue, resulting in Gaussian and Rician noise that make the image hard to read.

“Two main strategies are used for MRI denoising,” Zhang et al. wrote. “First, Rician data are treated directly, often in the spatial domain. Second, denoising is applied to the squared magnitude MR images, which follows a noncentral chi-square distribution with two freedom degrees whose noncentrality parameter is proportional to the underlying noise-free squared magnitude.”

Methods for denoising—namely the filtering, transform domain and statistical methods—are effective to a point, but your typical denoising algorithm designed for additive white Gaussian noise reduction probably won’t yield good results for Rician noise, the authors said. Even NeighShrink, an imaging denoising algorithm that’s been successful in reducing white Gaussian noise in past studies, struggles to make sense of Rician noise.

Zhang and co-authors combined the NeighShrink algorithm with CURE to determine NeighShrink’s optimal threshold, calling the product NeighShrinkCURE and combining it with fast bilateral filtering (BF) and fast cycle-spin technology to remove the noise in MR images while retaining image quality. Since NeighShrink’s denoising function relates only to wavelet coefficients, the team first developed a special wavelet domain CURE estimation tuned to the algorithm.

BF and fast cycle-spin technology were used to improve the performance of the algorithm at a minimal computational cost. The researchers found that when they applied NeighShrinkCURE to MR images artificially degraded by noise or downloaded from Brainweb, results were superior to two similar wavelet domain denoising algorithms (Iterative BF and LMMSE).

NeighShrinkCURE also demonstrated improved peak signal-to-noise ratio and structural similarity compared to other denoising methods. Still, its running time exceeds that of its competitors.

“Both quantitatively and qualitatively, the results show the efficiency of the proposed algorithm to MR image denoising,” Zhang et al. wrote. “Note that simple threshold shrinkage denoising methods optimized by CURE can also yield good results, however, the NeighShrink method makes use only of intra-scale dependencies between wavelet coefficients. Further denoising gains are likely to improve NeighShrink by introducing inter-scale dependencies.”