SD-OCT视网膜图像神经纤维层的检测与研究

发布时间：2018-03-14 23:18

本文选题：青光眼　切入点：光学相干断层技术　出处：《南京理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：青光眼是导致人类眼睛失明的第二大视网膜疾病。其主要导致杯盘比增大以及视觉神经头(ONH)附近视网膜神经纤维层(RNFL)厚度变薄。因此测量视网膜神经纤维层的厚度可以辅助诊断青光眼等疾病。目前,有许多类型的医疗设备可以获取视网膜结构图像。尤其频域光学相干断层成像技术(SD-OCT)最为突出,可以清晰地展示视网膜层组织结构。本文利用图像处理和机器学习相关知识分割视觉神经头SD-OCT视网膜图像的神经纤维层,对诊断青光眼具有很重要的临床和研究价值。本文的主要研究如下:(1)视网膜层结构组织错综复杂,某些视网膜疾病更会导致层结构发生改变,如老年性黄斑变性形成玻璃膜疣导致视网膜色素上皮层(RPE)层向上拱起;青光眼破坏视盘结构,导致视网膜神经纤维层厚度变薄等。因此传统的图像分割算法无法直接有效地分割视网膜层。第三章提出一种基于随机森林的OCT视网膜神经纤维层(RNFL)分割算法,算法主要分为三步:第一步是预处理阶段,包括图像去噪和获取感兴趣区域。通过比较-些主流的去噪算法的去噪效果,最终选择BM3D滤波算法。由于OCT视网膜图像内界膜(ILM)层与视网膜色素上皮层(RPE)层之外呈现不反光的黑色区域,因此将感兴趣区域限制在这两层之间;第二步是特征提取与训练过程;第三步利用第二步的训练模型分类其余特征,得到粗视网膜神经纤维层。最后使用多项式平滑分割曲线。(2)青光眼的产生通常伴随着多个明显特征,例如RNFL层缺失或者杯盘比变化等。因此评估这些特征的相关性,对研究青光眼的发病原理具有重要的意义。为了计算杯盘比等参数,需要精确分割视盘视杯。第四章详细介绍了一种基于SVM的块搜索算法。该算法分为两步:第一步利用视网膜RPE断点与视盘的关系,粗略分割出杯盘位置;第二步,使用块搜索算法在上一步的粗杯盘位置搜索最有可能的以RPE断点为中心的块,精确定位视盘视杯。最后计算RNFL层厚度与杯盘比等参数的相关性。(3)最后文章介绍一款OCT视网膜图像层分割和三维重建软件,该软件可以同时分割视网膜五层,对噪声和疾病具有较强的鲁棒性。
[Abstract]:Glaucoma is the second largest retinal disease leading to blindness in human eyes. It mainly results in the increase of the cup / disc ratio and the thinning of the retinal nerve fiber layer (RNFL) near the optic nerve head (ONH). Therefore, the thickness of the retinal nerve fiber layer is measured. Degree can assist in the diagnosis of glaucoma and other diseases. There are many types of medical devices that can capture retinal structure images, especially frequency-domain optical coherence tomography (SD-OCTs). This paper uses image processing and machine learning to segment the nerve fiber layer of visual nerve head SD-OCT retina image. In this paper, the main research is as follows: 1) the structure of the retinal layer is complicated, and some retinal diseases may lead to changes in the structure of the retinal layer. In the case of age-related macular degeneration, the formation of vitreous verruca causes the retinal pigment epithelium layer to arch upward, and glaucoma destroys the structure of the optic disc. The thickness of retinal nerve fiber layer is thinned and so on. So traditional image segmentation algorithm can not directly and effectively segment retinal layer. In chapter 3, a OCT retinal nerve fiber layer segmentation algorithm based on random forest is proposed. The algorithm is mainly divided into three steps: the first step is the pre-processing stage, including image denoising and obtaining the region of interest. Finally, the BM3D filtering algorithm is chosen. Because there is a dark region which is not reflected from the inner boundary layer and the retinal pigment epithelium layer, the region of interest is confined between the two layers. The second step is the feature extraction and training process, the third step classifies the other features using the training model of the second step. The formation of glaucoma is usually accompanied by several obvious features, such as the absence of RNFL layer or the change of cup / disc ratio, etc. Therefore, the correlation of these features is evaluated. It is of great significance to study the pathogenesis of glaucoma. In order to calculate the ratio of cup to disc, In Chapter 4th, a block search algorithm based on SVM is introduced in detail. The algorithm is divided into two steps: the first step is to roughly segment the position of the cup and disc by using the relationship between the retinal RPE breakpoint and the disc, and the second step is to segment the cup and disc position. The block search algorithm is used to search for the most likely block centered on the RPE breakpoint in the previous coarse cup and disk position. Finally, we calculate the correlation between the thickness of RNFL layer and the ratio of cup to disc. Finally, this paper introduces a software of OCT retinal image layer segmentation and 3D reconstruction, which can divide the five layers of retina simultaneously. It is robust to noise and disease.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R770.4;TP391.41

【相似文献】