基于脑部MRI的三维分割方法研究

发布时间：2018-02-27 01:19

本文关键词： 核磁共振成像隐马尔科夫模型模糊C均值模型空间相关性　出处：《吉林大学》2017年硕士论文　论文类型：学位论文

【摘要】：医学脑图像的三维处理是当前计算机视觉中的一个活跃的研究主题,它为修复手术、放射治疗、立体定向神经外科及其他医学领域提供了有意义的脑组织信息。脑组织各部分的分割对于定量分析具有极其重要的意义,这些定量测量是评估中枢神经系统疾病进展的关键因素,极大地提高了诊断和监测疾病演变的能力。核磁共振成像(Magnetic Resonance Imaging,MRI)技术目前被大规模应用在诊断人脑疾病等医疗领域,由于其具有成像清晰、对人体安全无创等优点,而被越来越多的人关注。然而,MRI图像易受噪声、偏移场、部分容积效应等因素的影响,使得缺乏医学先验知识的以计算机为基础的自动分割算法的性能严重下降。因此科学的脑组织分割方法仍是医学图像处理领域的重点研究课题。本文重点研究了基于核的模糊C均值算法(FuzzyC-means,FCM)及隐马尔科夫模型(Hidden Markov Models,HMMs)相结合的人脑MRI三维图像分割框架。通过多层脑切片的特征向量序列对HMMs模型进行训练,并利用训练模型对人脑MRI切片图像进行分类,再利用本文提出的改进的基于核的自适应正则化的FCM算法对每层脑切片图像进行分割,同时由切片分类时得到的信息减轻算法对初值的依赖,进而提高三维人脑组织分割的精确度。针对以上思想,本文具体的研究工作如下:1.对人脑切片图像进行去除颅骨组织,去噪等预处理,对切片进行预分割得到HMMs模型的观察序列,使用连续HMMs模型对每张切片进行分类,同时为后续的改进FCM算法提供更加丰富的信息。2.针对传统FCM算法和相关的改进算法中未充分考虑隶属度受空间关系影响的缺点,提出改进的基于核的自适应正则化的FCM算法。改进算法引入隶属度的空间约束函数,通过像素邻域内的隶属度信息调节像素对聚类中心的隶属程度,权衡空间信息对隶属度的影响,提高算法的聚类效果和抗噪性。同时利用已分割的前一帧的切片信息,进一步提升本文的基于核的自适应正则化的FCM算法对人脑切片分割的准确率,最终组合多层分割后的脑切片完成体分割。3.通过几组对比实验,验证了改进的基于核的自适应正则化FCM方法和基于HMMs模型的MRI人脑自适应切片分割算法的有效性。
[Abstract]:The 3D processing of medical brain image is an active research topic in computer vision, it is for repair surgery, radiotherapy, Stereotactic neurosurgery and other medical fields provide meaningful information on brain tissue. These quantitative measurements are key factors in assessing the progress of central nervous system diseases and greatly improve the ability to diagnose and monitor the evolution of diseases. Magnetic resonance imaging (MRI) Resonance imaging techniques are now widely used in medical fields such as the diagnosis of human brain diseases. Because of its advantages of clear imaging and non-invasive to human safety, MRI images are concerned by more and more people. However, MRI images are easily affected by noise, migration field, partial volume effect and so on. The performance of the computer-based automatic segmentation algorithm, which lacks medical prior knowledge, is seriously degraded. Therefore, the scientific brain tissue segmentation method is still a key research topic in the field of medical image processing. Based on kernel fuzzy C-means algorithm (FuzzyC-means-FCM) and Hidden Markov models (HMMs), a framework for human brain MRI 3D image segmentation is proposed. The HMMs model is trained by the feature vector sequence of multilayer brain slices. Using the training model to classify the human brain MRI slice images, and then using the improved kernel-based adaptive regularization FCM algorithm to segment each slice image. At the same time, the information obtained from the slice classification reduces the dependence of the algorithm on the initial value, thereby improving the accuracy of three-dimensional human brain tissue segmentation. In view of the above ideas, the specific research work of this paper is as follows: 1. Remove the skull tissue from the slice image of human brain. The HMMs model observation sequence is obtained by pre-segmentation of slice, and each slice is classified by continuous HMMs model. At the same time, it provides more abundant information for the following improved FCM algorithm. Aiming at the disadvantage of traditional FCM algorithm and related improved algorithm, the influence of membership degree on spatial relationship is not fully considered. An improved kernel-based adaptive regularization (FCM) algorithm is proposed, in which the spatial constraint function of membership degree is introduced, and the membership degree of pixel to cluster center is adjusted by membership information in pixel neighborhood. Considering the influence of spatial information on membership degree, the clustering effect and anti-noise of the algorithm are improved. At the same time, the slice information of the previous frame is used. Further improve the kernel based adaptive regularization of the FCM algorithm for human brain segmentation accuracy, the final combination of multi-layer segmentation of brain slices complete volume segmentation .3.Through several groups of comparative experiments, The improved kernel-based adaptive regularization FCM method and the MRI human brain adaptive slice segmentation algorithm based on HMMs model are validated.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R445.2;TP391.41

【相似文献】