徐晓玲,刘沂玲,刘且根,等.基于原始对偶字典学习的磁共振复数图像去噪[J].深圳大学学报理工版,2016,33(06):578-585.[doi:10.3724/SP.J.1249.2016.06578 ]
Xu Xiaoling,Liu Yiling,Liu Qiegen,et al.A novel PDL denoising algorithm for magnetic resonance complex images[J].Journal of Shenzhen University Science and Engineering,2016,33(06):578-585.[doi:10.3724/SP.J.1249.2016.06578 ]
基于原始对偶字典学习的磁共振复数图像去噪

南昌大学信息工程学院,江西南昌330031

图像处理; 字典学习; 对偶字典学习算法; 磁共振复数图像去噪; 莱斯分布; 核奇异值分解算法; 三维块匹配滤波算法

A novel PDL denoising algorithm for magnetic resonance complex images
Xu Xiaoling, Liu Yiling, Liu Qiegen, and Zhang Minghui

Xu Xiaoling, Liu Yiling, Liu Qiegen, and Zhang MinghuiSchool of Information Engineering, Nanchang University, Nanchang 330031, Jiangxi Province, P.R.China

image processing; dictionary learning; predual dictionary learning; magnetic resonance complex images denoising; Rician distribution; kernel singular value decomposition(K-SVD); block-matching and 3D filtering

DOI: 10.3724/SP.J.1249.2016.06578

备注

针对磁共振(magnetic resonance, MR)幅度图像中带有不易去除的与信号相关的莱斯(Rician)噪声问题,利用其复数图像中的实部与虚部所含噪声为不相关的加性高斯白噪声这一特性,代替对幅度图像直接去噪,提出将原始对偶字典学习(predual dictionary learning, PDL)算法用于对MR复数图像的实部与虚部分别进行去噪,然后组合得到幅度图像的方法.经仿真实验和在HT-MRSI50-50(50 mm)1.2 T小动物核磁共振系统中的实际应用,证明所提方法较直接对幅度图像去噪取得更好的效果,在有效去除MR图像噪声的同时能较好地保持图像中的细节.与经典的字典学习算法核奇异值分解(kernel singular value decomposition,K-SVD)相比,PDL算法去噪效果优于K-SVD算法,而运算速度提高约5倍. 与经典的基于非局部相似块的三维块匹配滤波(block-matching and 3D filtering, BM3D)算法相比,在噪声水平较低时PDL算法略优于BM3D算法,噪声水平较高时BM3D算法略优于PDL算法,两者总体比较接近.

The noise in magnetic resonance(MR)magnitude images presents a signal-dependent Rician distribution, which is quite difficult to remove. We propose a novel denoising approach to MR images. Since the noise in MR complex images' real and imaginary parts is additive and uncorrelated zero-mean Gaussian noise, we first apply the predual dictionary learning algorithm(PDL)to respectively denoise the real and imaginary parts of the MR complex images. We, then, combine the two parts to be a denoised MR magnitude image. Extensive simulation experimental results and practical applications on small animal MRI system HT-MRSI50-50(50 mm)1.2 T demonstrate that the proposed method is able to effectively remove noises while keeping the details of images. Compared with the classical dictionary learning algorithm of kernel singular value decomposition(K-SVD), the PDL algorithm not only attains better results but also requires less time, which is nearly five times faster than the K-SVD algorithm. Compared with block-matching and 3D filtering(BM3D), the proposed approach is slightly superior to BM3D in lower noise, and BM3D is slightly superior to ours in higher noise.

·