基于多图谱分割的融合算法研究

发布时间：2018-06-10 17:51

本文选题：人脑MR + 多图谱　；参考：《宁夏大学》2017年硕士论文

【摘要】：人体大脑结构复杂且功能各异,如海马体、扁桃体、颞上回、小脑、脑干、尾状核等关键脑结构与多种脑部疾病息息相关,对其精准分割是临床诊断中医生进行相关定量分析的前提,因此多图谱分割技术成为了当前国内外的研究重点。多图谱分割技术主要包括两个关键步骤,分别为图像配准和标记融合。将多个图谱与目标图像进行配准并选择合适的标记融合算法对配准后的图谱进行融合得到最终的分割结果。为了使得分割的结果更准确,需要选择合适的标记融合算法,以便于配准后的图像在融合过程中实现高精度,从而对每个初始分割中的信息进行有效的提取,使得最终的分割结果具有代表性。标记融合方法中用的比较广泛的有多数表决算法(Majority Voting,MV)[1]、STAPLE算法[2](Simultaneous Truth and Performance Level Estimation)和COLLATE算法[3](Consensus Level,Labeler accuracy and Truth Estimation)等。MV没有考虑到各个分割图像的差异性,STAPLE算法没有利用图像的先验信息。为了获得更高的分割精度,本文首先对脑部MR图像进行预处理,包括颅骨剔除、滤波、灰度归一化以及直方图匹配等处理并对多个组织进行配准,然后对基于配准的多图谱融合算法进行深入的研究并进行改进,主要内容如下:(1)围绕人脑MR图像,研究分析了当前使用比较广泛的MV融合算法和STAPLE融合算法,并使用这两种方法对配准后的脑部图像的多个组织进行融合,同时选择与金标准的相似性测度作为融合结果的评价标准,将这两种方法融合的结果与最优单图谱分割结果进行比较。(2)在MV融合算法的基础上提出一种新的加权改进融合算法(Weight-Voting),利用图谱和目标图像之间的相似性测度作为图像融合的权重,并分别对多个配准后的脑部组织进行融合,并将本文算法分别与最优单图谱、MV、STAPLE融合算法进行了比较。实验结果表明,多图谱分割方法分割精度要高于最优单图谱分割方法,本文提出的新融合改进算法性能优于最优单图谱、MV以及STAPLE融合算法,验证了本文提出的算法在医学图像分割方面的有效性和准确性。
[Abstract]:The complex and diverse structure of the human brain, such as the hippocampus, tonsils, superior temporal gyrus, cerebellum, brain stem, caudate nucleus, and other key brain structures are closely related to a variety of brain diseases. Accurate segmentation is the premise of quantitative analysis for doctors in clinical diagnosis, so multi-spectrum segmentation technology has become the focus of research at home and abroad. Multi-spectrum segmentation includes two key steps: image registration and label fusion. Finally, the final segmentation results are obtained by matching multiple maps with target images and selecting the appropriate label fusion algorithm. In order to make the segmentation result more accurate, it is necessary to select the appropriate label fusion algorithm, so that the registration image can achieve high accuracy in the fusion process, so that the information in each initial segmentation can be extracted effectively. The final segmentation results are representative. The majority voting algorithm (Majority VotingMV) [1] is a simple truth and performance level estimation algorithm [2] and a consensus level estimation algorithm [3]. MV does not take into account the difference of each segmented image and the prior information of the image. In order to achieve higher segmentation accuracy, the brain Mr image is preprocessed, including skull removal, filtering, gray normalization and histogram matching, and registration of multiple tissues is carried out. Then the multi-map fusion algorithm based on registration is deeply studied and improved. The main contents are as follows: 1) focusing on the human brain Mr image, the current widely used MV fusion algorithm and STAPLE fusion algorithm are studied and analyzed. The two methods are used to fuse multiple tissues of the brain image after registration, and the similarity measure with the gold standard is chosen as the evaluation criterion of the fusion results. Comparing the results of these two methods with the results of optimal single map segmentation, we propose a new weighted improved fusion algorithm, Weight-Votingn, based on the MV fusion algorithm. The similarity measure between the map and the target image is used as the measure of the similarity between the map and the target image. The weight of image fusion, The fusion of multiple brain tissues after registration was performed, and the proposed algorithm was compared with the optimal single map MVS-STAPLE fusion algorithm. The experimental results show that the segmentation accuracy of the multi-map segmentation method is higher than that of the optimal single-map segmentation method, and the performance of the improved fusion algorithm proposed in this paper is better than that of the optimal single-map MV and STAPLE fusion algorithms. The validity and accuracy of the proposed algorithm in medical image segmentation are verified.
【学位授予单位】：宁夏大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】