SIA OpenIR  > 机器人学研究室
核磁共振心脏图像的分割算法研究
其他题名Research on Segmentation Algorithm in Cardiac Magnetic Resonance Images
熊晶晶1,2
导师王振洲
分类号TP391.41
关键词核磁共振心脏图像 左心室分割 阈值分割 斜率差分布 图像滤波
索取号TP391.41/X68/2018
页数61页
学位专业模式识别与智能系统
学位名称硕士
2018-05-17
学位授予单位中国科学院沈阳自动化研究所
学位授予地点沈阳
作者部门机器人学研究室
摘要本文以基于短轴核磁共振心脏序列图像的左心室分割作为研究课题,从核磁共振心脏图像的预处理,基于斜率差分布(Slope Difference Distribution, SDD)的左心室分割算法,核磁共振心脏图像的后处理三个方面展开研究,旨在设计一个有效的分割算法实现左心室的分割。本文的主要工作如下:1. 进行了核磁共振心脏图像的预处理工作。从实验数据库Medical Image Computing and Computer Assisted Intervention(MICCAI)感兴趣区域提取、感兴趣区域特性分析以及感兴趣区域去噪三 方面展开。分别使用五种常见的滤波方法如均值滤波、中值滤波、高斯权重核滤波器、小波滤波以及傅里叶变换滤波对核磁共振心脏图像进行滤波处理,通过对比实验结果,选取了高斯权重核滤波器作为感兴趣区域图像去噪方法。2. 进行了基于SDD的左心室分割算法研究。SDD阈值分割方法主要包括三步:通过傅里叶滤波得到平滑灰度直方图;通过最小二乘模型得到斜率差分布;由阈值选取原则得到最佳的波谷点作为分割的阈值。着重研究了傅里叶滤波的参数(带宽)和最小二乘模型的参数(拟合点个数)对斜率差分布曲线的影响。基于SDD阈值选择方法的优越性在于SDD方法中斜率差分布的波谷(候选阈值)是可以调整的,通过不断校准带宽和拟合点个数,能让算法确定一个最佳的阈值点。最后,对不同斜率差分布曲线情况下阈值选取的原则展开了深入的研究。3. 进行了核磁共振心脏图像的后处理工作。通过选择最大的对象、生成最小凸多边形、FFT滤波实现边界平滑和膨胀四个步骤进行后处理,将获得的左心室二值化分割结果做优化。通过后处理手段,极大的改进了基于SDD方法分割出的初始轮廓的精度。4. 进行了基于Natural Image Quality Evaluator(NIQE)的感兴趣区域的无参考图像质量评估,并根据质量分数将感兴趣区域图像分成图像质量好和图像质量差两类。在预处理和后处理手段完全相同的条件下,进行基于SDD的阈值分割算法和其它15种国际上流行的分割算法的对比实验。实验数据表明SDD方法实现了1.6758毫米的平均垂直距离,0.8973的面积重叠率。与其他方法的比较表明,SDD方法在识别核磁共振心脏图像中的左心室边界最为准确。
其他摘要In this paper, the left ventricle segmentation is based on the short axis cardiac MR imaging. From the preprocessing of the MR images, the left ventricle segmentation algorithm based on Slope Difference Distribution (SDD) and the post-processing of the MR images are studied, aiming at designing an effective segmentation method to overcome the difficulty of left ventricular segmentation and realize the automatic segmentation of left ventricle. The main work of this article is as follows: 1. The preprocessing of MR images is performed. From the extraction of the region of interest (ROI) from the database named Medical Image Computing and Computer Assisted Intervention (MICCAI), the analysis of ROI and the ROI de-noising method. Five common filtering methods, such as mean filter, median filter, Gauss weight kernel filter, wavelet filter and Fourier transform filter, are used to filter the same ROI. By comparing the experimental results, the Gauss weighted kernel filter is selected as the ROI de-noising method. 2. The left ventricular segmentation algorithm based on SDD method is studied. The SDD threshold segmentation method mainly consists of three steps: smooth the original gray histogram of selected ROI by Fourier filtering; compute the slope difference distribution by the least square model; select the best threshold by the threshold selection principle. The influence of Fourier filter parameters (bandwidth) and least squares model parameters (number of fitting points) on the slope difference distribution curve is mainly studied. The superiority of the SDD threshold selection method is that the valleys (candidate threshold) of the slope difference distribution in SDD method can be adjusted by calibrating the bandwidth and the number of fitting points, which means that the SDD method can determine the best threshold. Finally, the principles of threshold selection under different slope difference distribution curves are thoroughly studied. 3. Post-processing of MRI cardiac images is studied. The original segmentation results of the left ventricle by SDD method are optimized by four steps: select the maximum object, generate convex hull image, boundary smoothing by FFT filtering and dilation of the boundary. The accuracy of segmentation based on SDD method is greatly improved by post-processing. 4. Non reference image quality evaluation of ROI based on Natural Image Quality Evaluator (NIQE) is applied, and the ROI images are divided into two categories (images with good image quality and images with poor image quality) according to the quality index. Under the condition of the same preprocessing and post processing, the comparison experiments between the SDD threshold selection method and the other 15 state-of-the-art segmentation algorithms are carried out. The experimental data show that the SDD method achieves 0.9246 DICE matrix and 1.3322 Average Perpendicular Distance (APD). The comparison with other methods shows that the SDD method is the most accurate way to identify the left ventricle in cardiac MR images.
语种中文
产权排序1
文献类型学位论文
条目标识符http://ir.sia.cn/handle/173321/21810
专题机器人学研究室
作者单位1.中国科学院沈阳自动化研究所
2.中国科学院大学
推荐引用方式
GB/T 7714
熊晶晶. 核磁共振心脏图像的分割算法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2018.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
核磁共振心脏图像的分割算法研究.pdf(19054KB)学位论文 开放获取CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[熊晶晶]的文章
百度学术
百度学术中相似的文章
[熊晶晶]的文章
必应学术
必应学术中相似的文章
[熊晶晶]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。