基于方向尺度描述子与稀疏编码的海马体分割
发布时间：2018-08-11 18:46
【摘要】：海马体病变与神经疾病息息相关,海马体解剖结构的不规则性以及与周围组织结构如杏仁体边界模糊增加了分割海马体的难度。目前较流行的图像分割算法较适用于分割规则器官或大器官,而海马体体积较小,形状不规则,因此常用的图像分割算法不能达到理想的分割精度。而在常用的基于图谱的分割算法中多以基于灰度的描述子描述图像特征,基于灰度的图像特征在描述亮暗不均匀图谱时辨识度差,本论文提出一种新的识别度较高的图像特征描述子——方向-尺度描述子(orientation-scale descriptor OSD),然后结合稀疏编码算法提出一种新的基于方向尺度描述子和稀疏编码(orientation-scale descriptor and spare coding OSDSC)海马体分割算法,提高海马体分割精度。不同于主流的基于字典学习的方法,OSDSC算法用同时包含灰度纹理信息和空间结构信息的方向-尺度描述子(orientation-scale descriptor OSD)代替低维特征来描述像素特征,OSD的优点是它同时包含多种低维特征且能降低图谱间灰度不均匀性的影响。OSDSC算法包括四个步骤:首先,图像预处理。第二,特征提取:提取待分割图像像素和图谱图像像素的方向-尺度描述子。第三,字典构建及稀疏编码:选取图谱像素的方向-尺度描述子为目标像素构建特有字典,用特有字典近似表达即重建目标像素并得到稀疏编码系数;第四,标号融合及阈值判定。融合图谱像素的标号和编码系数得到目标像素的标号估计值;阈值判定估计值完成分割。为了验证OSDSC算法分割的准确性,分别用OSDSC算法,Simple,Major Voting,Staple,Collate算法分割MICCAI数据库中海马体,以Dice值作为分割评判标准,实验结果表明OSDSC方向-尺度描述子的分割精度高于Simple,Major Voting,Staple,Collate 算法。
[Abstract]:Hippocampal lesions are closely related to neurological diseases. The irregularity of hippocampal anatomical structure and the blurring of the surrounding tissue such as amygdala increase the difficulty of hippocampal body segmentation. At present, the popular image segmentation algorithms are more suitable for regular organs or large organs, but the hippocampal body is small in volume and irregular in shape, so the commonly used image segmentation algorithms can not achieve the ideal segmentation accuracy. However, the image features are often described by grayscale based descriptors, and the recognition degree of grayscale based image features is poor when describing the non-uniform spectrum of light and dark. In this paper, a new image feature descriptor with high recognition, direction-scale descriptor (orientation-scale descriptor OSD), and a new directional scale descriptor and sparse coding (orientation-scale descriptor and spare coding) based on sparse coding algorithm are proposed. OSDSC) Hippocampal Segmentation algorithm, Improve the accuracy of hippocampal segmentation. Different from the mainstream Dictionary Learning-based approach, OSD algorithm uses direction-scale descriptors that contain both grayscale texture information and spatial structure information (orientation-scale descriptor OSD) instead of low-dimensional features to describe pixel features) the advantage of OSD is that it simultaneously packets. The OSDSC algorithm includes four steps: firstly, it can reduce the influence of grayscale heterogeneity between maps. Image preprocessing. Second, feature extraction: extracting the direction-scale descriptor of pixels of image to be segmented and pixels of atlas image. Thirdly, dictionary construction and sparse coding: selecting the direction-scale descriptor of map pixels as target pixels to construct special dictionaries, using the approximate representation of unique dictionaries to reconstruct target pixels and obtain sparse coding coefficients; fourth, Label fusion and threshold determination. The label estimation value of the target pixel is obtained by combining the labeling and coding coefficients of the map pixels, and the threshold decision estimation value is segmented. In order to verify the accuracy of OSDSC segmentation, the Dice algorithm is used to segment the hippocampus in MICCAI database using OSDSC algorithm. The experimental results show that the segmentation accuracy of OSDSC direction-scale descriptor is higher than that of simple Major Votingling StapleCollate algorithm.
【学位授予单位】：南方医科大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
本文编号：2177891
[Abstract]:Hippocampal lesions are closely related to neurological diseases. The irregularity of hippocampal anatomical structure and the blurring of the surrounding tissue such as amygdala increase the difficulty of hippocampal body segmentation. At present, the popular image segmentation algorithms are more suitable for regular organs or large organs, but the hippocampal body is small in volume and irregular in shape, so the commonly used image segmentation algorithms can not achieve the ideal segmentation accuracy. However, the image features are often described by grayscale based descriptors, and the recognition degree of grayscale based image features is poor when describing the non-uniform spectrum of light and dark. In this paper, a new image feature descriptor with high recognition, direction-scale descriptor (orientation-scale descriptor OSD), and a new directional scale descriptor and sparse coding (orientation-scale descriptor and spare coding) based on sparse coding algorithm are proposed. OSDSC) Hippocampal Segmentation algorithm, Improve the accuracy of hippocampal segmentation. Different from the mainstream Dictionary Learning-based approach, OSD algorithm uses direction-scale descriptors that contain both grayscale texture information and spatial structure information (orientation-scale descriptor OSD) instead of low-dimensional features to describe pixel features) the advantage of OSD is that it simultaneously packets. The OSDSC algorithm includes four steps: firstly, it can reduce the influence of grayscale heterogeneity between maps. Image preprocessing. Second, feature extraction: extracting the direction-scale descriptor of pixels of image to be segmented and pixels of atlas image. Thirdly, dictionary construction and sparse coding: selecting the direction-scale descriptor of map pixels as target pixels to construct special dictionaries, using the approximate representation of unique dictionaries to reconstruct target pixels and obtain sparse coding coefficients; fourth, Label fusion and threshold determination. The label estimation value of the target pixel is obtained by combining the labeling and coding coefficients of the map pixels, and the threshold decision estimation value is segmented. In order to verify the accuracy of OSDSC segmentation, the Dice algorithm is used to segment the hippocampus in MICCAI database using OSDSC algorithm. The experimental results show that the segmentation accuracy of OSDSC direction-scale descriptor is higher than that of simple Major Votingling StapleCollate algorithm.
【学位授予单位】：南方医科大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
【参考文献】
相关期刊论文 前1条
1 杨晖;;图像分割的阈值法研究[J];辽宁大学学报(自然科学版);2006年02期
，本文编号：2177891
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2177891.html
