基于字典学习与等效视数的低剂量CT伪影抑制算法  

Low-Dose CT Artifact Suppression Algorithm Based on Dictionary Learning and Equivalent Number of Looks

作　　者：杨一鸣 刘祎[1] 桂志国[1] YANG Yi-ming;LIU Yi;GUI Zhi-guo(North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学生物医学成像与影像大数据重点实验室

出　　处：《中北大学学报（自然科学版）》2019年第6期559-567,共9页Journal of North University of China(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61671413);国家重点研发计划资助项目(2016YFC0101602)

摘　　要：针对低剂量CT图像出现条形伪影的现象,提出了一种基于字典学习与等效视数(ENL)的伪影抑制算法.该方法首先利用平稳小波变换(SWT)对低剂量CT图像进行单层分解,并对高频图像训练字典,然后利用等效视数(ENL)对字典进行分区得到伪影字典和特征字典,并只对特征原子进行稀疏编码,经小波逆变换(ISWT)后得到处理的CT图像;然后,采用双边滤波器对处理后的CT图像进行分解并训练高频字典,通过判断等效视数(ENL)来摒弃伪影字典,从而去除高频图像残留的伪影和噪声,达到抑制条形伪影的目的.实验结果表明,与总变分降噪算法、K-奇异值分解(K-SVD)算法和三维块匹配滤波(BM3D)算法对比,该算法在抑制条形伪影的同时保留了更多的边缘和细节信息,并具有较高的结构相似性和峰值信噪比.An artifact suppression algorithm based on dictionary learning and equivalent number of looks(ENL)was proposed for low-dose CT images with streak artifacts.The method firstly used a stationary wavelet transform(SWT)to perform single-layer decomposition of low-dose CT images,and trained the dictionary for high frequency images.Then used the equivalent number of looks(ENL)to partition the dictionary to get the artifact dictionary and feature dictionary.And only the atom from feature dictionary was sparsely encoded,and the processed CT image was obtained after the inverse stationary wavelet transform(ISWT).Then,the processed CT image was decomposed and the high frequency dictionary was trained by using a bilateral filter,and the artifact dictionary was discarded by judging the equivalent number of looks(ENL).Thereby,residual artifacts and noises remaining in the high frequency image were removed,and the purpose of suppressing streak artifacts was achieved.Compared with the total variation noise reduction algorithm,K-singular value decomposition(K-SVD)algorithm and blockmatching 3-D(BM3 D)filter algorithm,the experimental results show that the proposed algorithm preserves more edge and detail information while suppressing streak artifacts,and has higher structural similarity and peak signal-to-noise ratio.

关 键 词：低剂量CT 字典学习 稀疏表示 等效视数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]