糖尿病性视网膜病变眼底图像微脉瘤检测

发布时间：2018-12-06 14:11

【摘要】：随着社会的发展，人们生活水平的提高，在我国糖尿病发病率呈逐年上升趋势，糖尿病性视网膜病变是糖尿病的并发症之一，也是导致失明或视觉损伤的主要病因之一，如果能够得到尽早的诊断和治疗，则可以有效的控制病情的发展。临床上，眼底图像是眼科疾病诊断的重要依据，眼底图像定期检查已成为糖尿病视网膜病变的重要筛选手段。作为最早出现和最小能被观测的病变对象，微脉瘤的检测和定位对于早期病变的防治显得尤为重要。由于背景纹理及噪声、视盘、黄斑、血管等因素的干扰，眼底图像普遍存在的非均匀照明和非均匀对比度现象，微脉瘤本身的尺寸差异以及因对比度弱而难以观测等因素的影响，相关病变区域定位和分割很容易出现错误，因此微脉瘤的检测目前仍是一个比较困难的任务。已有文献报道中，关于弱微脉瘤的检测准确率很低，同时微脉瘤的检测易受血管提取的影响，据观察有30%的误否认和90%误确认结果和不准确、不完整的血管结构提取有关，因此如何提高正判率同时降低误判率仍将研究人员致力解决的问题。本文以数字彩色眼底图像处理与识别为基础，专门针对微脉瘤的检测做了较深入的研究，致力于提高微脉瘤检测的鲁棒性和准确性，主要工作包括以下内容：从背景估计建模的角度提出一种基于Mahalanobis距离的疑似微脉瘤判别方法，由于血管是造成大量误确认结果的主要因素，因此为获得尽可能准确完整的血管结构以便排除血管像素，我们提出了一种基于Gabor滤波器的非血管结构抑制算子并结合多尺度和多滞后阈值技术的精细血管骨架提取方法，并在后期处理中，我们采用相对简单的形状分析和双环滤波进一步去除了虚假病变点。实验结果表明，，该方法性能优于或逼近于其它同类方法，而且能够极大提高邻近血管的微脉瘤检测精度。研究了一种结合多尺度高斯匹配滤波和集成分类方法的微脉瘤检测方法。由于微脉瘤在空间分布符合二维高斯分布特性，同时存在较大的尺寸差异变化，通过多尺度高斯匹配滤波，首先筛选出疑似微脉瘤对象，并作为种子点利用区域生长技术分割出病变区域，进而提取病变区域特征信息，最终采用Adaboost神经网络集成分类器检测真实的微脉瘤病变。该方法在公开的ROC数据集上进行了测试，实验表明，检测性能优于以往的双环滤波和形态学方法。
[Abstract]:With the development of society and the improvement of people's living standard, the incidence of diabetes is increasing year by year in our country. Diabetic retinopathy is one of the complications of diabetes, and it is also one of the main causes of blindness or visual impairment. If you can get early diagnosis and treatment, you can effectively control the development of the disease. In clinic, fundus image is an important basis for the diagnosis of ophthalmic diseases. Regular examination of fundus image has become an important screening method for diabetic retinopathy. As the earliest and least observable lesions, the detection and localization of microvein tumors is particularly important for the prevention and treatment of early lesions. Because of background texture and noise, visual disc, macula, blood vessel and other factors interference, there are non-uniform illumination and non-uniform contrast phenomenon in the fundus image. Because of the difference in the size of microvein tumor itself and the influence of some factors such as weak contrast, the location and segmentation of the related lesions are prone to errors, so the detection of microvein tumor is still a difficult task at present. It has been reported that the detection accuracy of weak microvein tumor is very low, and the detection of microvein tumor is easy to be affected by blood vessel extraction. It is observed that 30% false denial and 90% false confirmation result and inaccurate, incomplete vascular structure extraction are related to the detection of microvein tumor. Therefore, how to improve the positive judgment rate and reduce the false judgment rate will be solved by researchers. Based on the digital color fundus image processing and recognition, this paper makes a deep research on the detection of microvein tumor, and devotes to improving the robustness and accuracy of the detection of microvein tumor. The main work includes the following: from the perspective of background estimation modeling, a method for identifying suspected microvascular tumors based on Mahalanobis distance is proposed, because blood vessels are the main factors causing a large number of false confirmation results. Therefore, in order to obtain as accurate and complete a vascular structure as possible in order to exclude the pixels of the vessel, we propose a method of extracting the fine vascular skeleton based on the Gabor filter and combining the multi-scale and multi-delay threshold techniques. In the post-processing, we further remove the false lesions by using relatively simple shape analysis and double-loop filtering. The experimental results show that the performance of this method is superior to or close to that of other similar methods, and it can greatly improve the detection accuracy of microvein tumors in adjacent blood vessels. A multi-scale Gao Si matched filtering and integrated classification method for microvein tumor detection is studied. Because the spatial distribution of microveinoma accords with the two-dimensional Gao Si distribution and there is a great difference in size, the suspected microveinoma objects are first screened by the multi-scale Gao Si matched filter. As a seed point, the region growing technique is used to segment the lesion area, and then extract the characteristic information of the lesion area. Finally, the Adaboost neural network integrated classifier is used to detect the true microvein tumor lesion. The method is tested on the open ROC dataset. The experimental results show that the detection performance is better than that of the previous dual-loop filtering and morphology methods.
【学位授予单位】：湘潭大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：R587.2;R774.1;TP391.41

【参考文献】