低剂量CT图像重建与伪影校正算法研究

发布时间：2018-04-23 14:48

本文选题：低剂量CT + 图像重建　；参考：《东北大学》2015年博士论文

【摘要】：CT是疾病诊疗过程中常用的设备之一,在疾病的诊断中无法用其它设备取代CT。但作为一种放射型设备,CT的广泛使用增加了患者吸收的X射线剂量,对身体健康造成潜在的危害。由于低剂量CT可以减少X射线对患者的伤害,因此在不影响重建图像质量的前提下,如何降低X射线剂量受到越来越多的关注和研究。通常,可以通过减少扫描角度、减少扫描时间和降低放射电流三种方式来降低放射的剂量。扫描角度的减少会带来投影数据的不完备,通过迭代重建算法可以很好的解决这种数据不完备问题；扫描时间的减少或放射电流的降低使得光子数减少,从而引起投影数据的噪声过大,进而导致重建后的图像含有严重的高斯噪声,基于图像域和投影域的去噪方法可以有效的抑制噪声。此外,对带有金属植入体的患者进行扫描时,重建后的CT图像内部会产生严重的金属伪影,如何精确地恢复出被金属影响的投影数据,是一个很有意义的问题。最后探讨的是因探测器单元的异常而引起图像中产生环状伪影的问题,通过投影域滤波的方法能够有效地去除环状伪影。因此,本论文主要围绕低剂量CT的图像重建和伪影的校正过程中的关键点和难点进行了如下的研究：(1)针对如何从不完备的投影数据重建出高质量CT图像的问题,提出了一种有序子集重建算法。该算法利用投影到凸集合加快重建的收敛速率,结合全变分最小化和快速一阶方法来减少重建的迭代次数,采用分裂Bregman交替方向法求解优化问题。通过对仿真数据和实际数据进行实验,验证了提出的方法在不完备投影数据情况下的重建效果,为算法的实际应用提供了理论和实验基础。(2)考虑CT图像的高斯噪声问题,提出了一种基于CUDA加速的三维全变分最小化算法。该算法自动计算迭代步长,利用梯度下降法减少图像的全变分。通过对仿真和实际数据的实验结果进行分析,验证了提出的加速方法在保持图像的纹理和边界信息的前提下,可以有效地降低计算时间,减少大部分的高斯噪声。(3)针对抑制投影数据低剂量噪声和去噪模型的参数选择问题,根据投影域去噪的阶段不同,分别提出了基于对数前和对数后的四种去噪算法。在对数前的投影去噪中,提出基于一维投影升维的自适应步长梯度下降法、基于二维投影数据的自适应步长梯度下降法和自动计算惩罚参数的乘子交替方向法。在对数后的投影数据去噪过程中,针对对数后的噪声模型,提出一种自动计算噪声方差的去噪算法。该算法利用方差稳定性变换将信号依赖的高斯噪声转换为独立高斯噪声,通过噪声方差评估的方法确定去噪参数,采用三维滤波块匹配法和基于参考图像的非局部均值法进行去噪。通过仿真和实际数据的实验,验证了本论文提出的四个算法的可行性和有效性。(4)为了去除图像中的金属伪影,提出了一种基于冗余表达的校正方法。首先,分割出CT图像中的金属区域,将其投影得到金属投影数据；然后,从原始投影数据中去除金属部分的投影数据；最后,利用冗余表达的算法恢复金属部分的投影数据,经过滤波反投影后,重建出校正后的图像。校正结果显示,该算法可以有效地减少重建图像的金属伪影。(5)为了降低图像中的环状伪影,提出了一种基于CUDA加速的环状伪影校正算法。该算法利用GPU加速二维投影数据的中值滤波过程。校正结果显示,与CPU处理速度相比,提出的算法在处理时间上明显更快,在消除了环状伪影的同时,保持了图像的空间分辨率。
[Abstract]:CT is one of the commonly used equipment in the process of diagnosis and treatment of disease. CT. can not be replaced by other equipment in the diagnosis of disease, but as a radioactive device, the extensive use of CT increases the dose of X ray absorbed by the patient, causing potential harm to health. Because low dose CT can reduce the harm of X ray to the patient, so it does not affect the weight of the patient. On the premise of building the image quality, how to reduce the dose of X ray is paid more and more attention and research. Usually, the dose of radiation can be reduced by reducing the scanning angle, reducing the scanning time and reducing the radiation current. The decrease of the scanning angle will bring the incomplete projection data, and the iterative reconstruction algorithm can be very good by the iterative reconstruction algorithm. To solve the problem of incomplete data, the reduction of scanning time or the decrease of the radiate current makes the number of photons reduce, which causes the noise of the projection data to be too large, and then the reconstructed image contains serious Gauss noise. The denoising method based on the image domain and the projection domain can effectively suppress the noise. When the patient is scanned, a serious metal artifact can be produced inside the reconstructed CT image. How to accurately restore the projected data affected by the metal is a very meaningful problem. Finally, the problem of the annular artifact in the image is caused by the anomaly of the detector unit, which can be filtered by the projection domain. Therefore, this paper focuses on the key and difficult points in the process of image reconstruction and artifact correction in low dose CT. (1) an ordered subset reconstruction algorithm is proposed for the problem of how to reconstruct high quality CT images from incomplete projection data. The convex set speeds up the convergence rate of the reconstruction, combines the total variation minimization and the fast first order method to reduce the iteration number of the reconstruction. The split Bregman alternating direction method is used to solve the optimization problem. The simulation data and the actual data are tested to verify the reconstruction effect of the proposed method in the case of not finished projection data, which is an algorithm. The practical application provides the theoretical and experimental basis. (2) considering the Gauss noise problem of CT images, a three dimensional total variational minimization algorithm based on CUDA acceleration is proposed. The algorithm automatically calculates the iteration step and reduces the total variation of the image by the gradient descent method. The results of the simulation and actual data are analyzed and verified. The proposed acceleration method can effectively reduce the computation time and reduce most of the Gauss noise on the premise of maintaining the texture and boundary information of the image. (3) according to the parameter selection of low dose noise and de-noising model for suppressing the projection data, according to the difference of the order of the denoising in the projection domain, the method is proposed based on the logarithm before and after the logarithm respectively. Four denoising algorithms. In the projection denoising before the logarithm, the adaptive step gradient descent method based on one dimension projection ascending dimension is proposed, the adaptive step gradient descent method based on the two-dimensional projection data and the multiplier alternate direction method are used to automatically calculate the penalty parameters. In the process of logarithmic projection data denoising, the logarithmic noise model is used. A denoising algorithm for automatic calculation of noise variance is proposed. The algorithm uses variance stability transform to transform the signal dependent Gauss noise into independent Gauss noise. The denoising parameters are determined by the method of noise variance evaluation. The denoising is carried out by the three-dimensional filter block matching method and the non local mean method based on the reference image. The experiment of actual data proves the feasibility and effectiveness of the four algorithms proposed in this paper. (4) in order to remove the metal artifacts in the image, a correction method based on redundant expression is proposed. First, the metal region in the CT image is separated and projected to the metal projection data; then, gold is removed from the original projection data. The projection data belongs to the part; finally, the projected data of the metal part is restored by the redundant expression algorithm. After the filtered back projection, the corrected image is reconstructed. The correction results show that the algorithm can effectively reduce the metal artifacts of the reconstructed image. (5) to reduce the ring artifact in the image, a CUDA acceleration based on the algorithm is proposed. The algorithm uses GPU to accelerate the median filtering process of the two-dimensional projection data. The correction results show that compared with the CPU processing speed, the proposed algorithm is significantly faster in processing time, while maintaining the spatial resolution of the image while eliminating the ring artifact.

【学位授予单位】：东北大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：R814.42;TP391.41

【参考文献】