基于多图谱的MR前列腺图像分割算法研究

发布时间：2018-06-04 23:17

本文选题：前列腺分割 + MR　；参考：《南方医科大学》2017年硕士论文

【摘要】：前列腺炎、前列腺增生、前列腺癌等疾病在男性中越来越普遍,前列腺癌已经是全球范围内男性第二位最常见的癌症。磁共振(Magnetic Resonance,MR)成像技术能够较好的显示前列腺内在组织结构,对于前列腺疾病的分析与诊断具有重要的临床意义。临床上,前列腺的大小、形状、相对周围组织器官的位置信息是对前列腺疾病及其病理阶段进行诊断和分析的重要前提,同时在前列腺切除术、放射治疗中也起着关键性的指导作用。因此,准确分割前列腺是至关重要的。但在MR图像中,由于成像技术的限制、前列腺结构本身的复杂性以及不同个体之间前列腺形状、大小和纹理信息的差异性,前列腺分割一直是近些年的一个难点。目前基于计算机的MR前列腺图像全自动分割算法分割精度不高,尚不能满足临床要求。目前比较常见的MR前列腺图像分割算法主要有基于分类器的方法、基于参数形变模型的方法和基于多图谱的分割方法等。基于分类器的分割算法很大程度上依赖于分类器和提取的特征的性能,且当数据分布不平衡时,容易导致分类精度急剧下降;基于形变模型的分割算法对初始轮廓的形状和位置敏感,常易陷入局部极值,算法的鲁棒性差、抗干扰性差;而基于多图谱的分割方法利用手动分割精度高的优势,将分割问题转化为配准问题。通过配准技术,它能有效的将手动勾画好的前列腺图谱的形状先验知识融入分割过程,很好的重现分割结果。因此,基于多图谱的分割方法是近几年来非常受关注的一种方法。基于多图谱的分割算法主要包含三个步骤:配准、图谱选择和图谱融合,其中,图谱选择和融合策略可以在一定程度上降低配准误差对分割过程所带来的影响,有效提高分割精度,是近些年来国内外学者研究的两个关键点。本文首先对目前比较经典的图谱融合策略进行了学习和研究,并在此基础上提出了一种距离场融合算法。距离场融合算法不再是对标号图像进行融合,而是对每个标号图像对应的距离场图像进行融合。与图谱标号相比,距离场不但反映了图像中任意像素点的标号信息,同时也包含了该像素点与目标边界的相对位置关系。本文提出的距离场融合算法主要基于以下两个假设:(1)假设MR样本与其对应的距离场(DF)样本分别位于两个非线性流形上,每个样本都可由其局部邻域内的样本线性表示;(2)在局部空间内,MR样本与其对应的DF样本之间近似为微分同胚映射。基于这两个假设,则可以把距离场融合过程中权重系数的求解问题转化为用MR字典样本来线性表示测试样本时字典系数的求解问题。本文采用测试样本自身邻域内的字典样本对其进行线性表示,并利用强调局部性的局部锚点嵌入算法(LAE)对字典系数进行求解,最后对相应的DF字典样本进行线性组合来获得每个测试样本所对应的距离场,再经过加权平均和阈值处理,从而得到最终的分割结果。实验中与目前国际上比较流行的标号融合算法(Major Voting、SIMPLE、STAPLE、Nonlocal Patch-based Label Fusion等)进行了对比,实验结果验证了距离场融合算法的有效性;其次,针对配准误差过大导致的分割效果较差的情况,本文引入了一种椭球形状先验,将多图谱分割与椭球形状先验相结合,提出了一种椭球先验约束的多图谱MR前列腺分割算法。椭球先验的引入,可以只针对椭球先验约束下的前列腺感兴趣区域进行图谱选择,避免前列腺周围组织与器官对图谱选择造成的干扰。同时,在图谱融合过程中加入椭球先验项进行约束,可以对通过配准技术引入的前列腺图谱形状先验进行校正和补偿,避免由配准误差引起的错误分割的情况。对50例MR前列腺图像进行分割实验,实验结果表明该算法对前列腺数据的分割结果Dice相似度均在80%以上,平均Dice相似度提高到了 88.27%。
[Abstract]:Prostatitis, prostatic hyperplasia and prostate cancer are becoming more and more common in men. Prostate cancer is the second most common cancer in the world. Magnetic Resonance (MR) imaging technology can better display the internal structure of the prostate, which is important for the analysis and diagnosis of prostate disease. Clinically, the size, shape, and location of the prostate is an important prerequisite for the diagnosis and analysis of the prostate disease and its pathological stage, and it also plays a key role in the prostatectomy and radiation therapy. Therefore, it is very important to segment the prostate accurately. However, the MR image is very important. Because of the limitation of imaging technology, the complexity of prostate structure itself and the difference of prostate shape, size and texture information between different individuals, the segmentation of the prostate has been a difficult point in recent years. At present, the accuracy of the automatic segmentation of MR prostate image based on computer is not high, and the clinical requirements can not be met. The common MR prostate image segmentation algorithms are mainly based on classifier based methods, based on parameter deformation model and multi map based segmentation. The segmentation algorithm based on classifier depends largely on the performance of classifier and extracted features, and it is easy to lead to classification precision when the data distribution is unbalanced. The segmentation algorithm based on the deformation model is sensitive to the shape and position of the initial contour, is often prone to fall into the local extremum, the algorithm is poor in robustness and poor in anti-interference, and the segmentation method based on multi map uses the advantage of high precision of manual segmentation to transform the segmentation problem into registration problem. By registration technology, it can effectively hand the hand The shape prior knowledge of the shape of the prostate map is integrated into the segmentation process, and the segmentation results are reproduced well. Therefore, the segmentation method based on multi map is a very popular method in recent years. The segmentation algorithm based on multi map mainly contains three steps: registration, graph selection and map fusion, among which, map selection and fusion The strategy can reduce the effect of registration error to the segmentation process to a certain extent and improve the segmentation accuracy effectively. It is the two key point of scholars at home and abroad in recent years. Firstly, this paper studies and studies the classical map fusion strategy. On this basis, a distance field fusion algorithm is proposed. The field fusion algorithm is no longer the fusion of the label image, but the fusion of the distance field images corresponding to each label image. Compared with the map label, the distance field not only reflects the label information of any pixel in the image, but also contains the relative position relation between the pixel points and the boundary of the target. The algorithm is based on the following two hypotheses: (1) assuming that the MR sample and its corresponding distance field (DF) samples are located on two nonlinear manifolds, each sample can be linearly represented by the sample in its local neighborhood; (2) in the local space, the MR sample and its corresponding DF sample are approximately differential homeomorphic mapping. Based on these two hypotheses, The problem of solving the coefficient of the weight coefficient in the distance field fusion process can be transformed into the dictionary coefficient for the linear representation of the test sample with the MR dictionary sample. This paper uses the dictionary sample in the test sample's own neighborhood to express it linearly, and uses the local anchored point embedding algorithm (LAE) to the dictionary coefficient. Finally, the corresponding DF dictionary samples are linearly combined to obtain the distance field corresponding to each test sample, and then the weighted average and threshold processing is used to get the final segmentation results. In the experiment, the Major Voting, SIMPLE, STAPLE, Nonlocal Patch-based Label F are more popular in the experiment. Usion and so on, the experimental results verify the effectiveness of the distance field fusion algorithm. Secondly, in the case of poor registration error caused by too large registration error, this paper introduces a type of ellipsoid shape prior, combining the multi map segmentation with the ellipsoid shape prior, and proposes a multi map MR prostate segmentation with ellipsoid prior constraints. The introduction of ellipsoid prior can only select the map of the region of interest of the prostate under the ellipsoid prior constraint, avoid the interference of the tissues and organs around the prostate to the selection of the atlas. At the same time, the ellipsoid prior item is added to the map fusion process to restrict the prostate atlas, which is introduced by the registration technique. The shape prior is corrected and compensated to avoid the error segmentation caused by registration error. 50 cases of MR prostate images are segmented. The experimental results show that the Dice similarity of the proposed algorithm is more than 80% for the segmentation results of the prostate data, and the average Dice similarity is increased to 88.27%.
【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】：R737.25;R445.2;TP391.41

【参考文献】