一种自适应低剂量CT图像质量改善算法-《郑州大学学报(工学版)》

文章信息/Info

作者:: 蒋慧琴; 徐玉凤; 马岭; 杨晓鹏; Toshiya Nakaguchi; 1.郑州大学信息工程学院,河南郑州,450001;2.郑州大学第一附属医院医学装备部,河南郑州,450052;3.日本千叶大学先端医工学研究中心, Chiba

Author(s):: Jiang Huiqin1; Xu Yufeng1; Maling1; Yang Xiaopeng2; Toshiya Nakaguchi3; 1. School of Information Engineering, Zhengzhou University, Zhengzhou, Henan, 450001; 2. Department of Medical Equipment, The First Affiliated Hospital of Zhengzhou University, Zhengzhou, Henan, 450052; 3. Advanced Medical Engineering Research Center, Chiba University, Chiba, Japan

Keywords:: Low-dose CT; Shearlet Transform; Quantum Noise; Bayesian Estimation; Noise Variance

摘要:: 针对低剂量CT(Low-dose CT，LDCT)扫描惠导致图像质量劣化问题，提出一种基于剪切波变换的低剂量CT图像质量改善算法。首先，利用Anscombe变换，将LDCT图像中的X射线量子噪声转化为近似服从Gaussian分布的噪声；其次，将变换后的LDCT图像转换成剪切波变换域并针对剪切波域上的低信噪比高频系数子带，利用剩余自相关功率改进噪声方差的计算精度并结合贝叶斯最大后验估计提取非噪声高频系数；最后，利用剪切波逆变换和Anscombe逆变换获得重构图像。大量的实验结果表明，该算法优于小波域的算法。其重构图像质量与基于小波域的算法相比，峰值信噪比（PSNR）平均提高52.2%，平均结构相似度（MSSIM）提高34.9%。

Abstract:: The low-does CT（LDCT）scanning is an effective way to reduce the X-ray radiation dose. However, quantum noise caused by the reduction of radiation dose leads to degradation of image quality. We proposed a quality improvement algorithm of low-dose CT images based on the shearlet transformation. Firstly, LDCT image was transformed using the Anscombe transform, and the quantum noise was transformed into noise which approx imately obeyed Gaussian distribution. Secondly, the transformed image is decomposed into low-frequency coefficient sub-bands and multi-directional high-frequency coefficient sub-bands based on shearlet transform. Then, for high-frequency coefficient sub-bands of the low signal noise ratio, a noise variance estimation method basedon the residual autocorrelation power (RAP) was proposed, which was combined with Bayesian maximumposterior probability method to obtain the more accurate non-noise high-frequency cofficients. Finally, the reconstructed image was obtained using the shearlet inverse transform and anscombe inverse transform. A series of experimental results of quantitative evakuation and visual effects showed that the proposed algorithm outperformed the de-noising method based on wavelet domain. The quality of the reconstructed image, compared with the denoising algorithm based on wavelet domain, the Peak Signal Noise Ratio(PSNR) was increased averagely by 52.2%,and the Mean Structure Similarity(MSSIM) was increased by 34.4%.