糖尿病视网膜病变的系统血管分析算法与辅助诊断

The prevalence of diabetes has increased significantly over the past decades and is now reaching epidemic proportions in China. Diabetic retinopathy (DR) is one of the main complications of diabetes. The onset of diabetic retinopathy greatly reduces the quality of life and eventually leads to blindness. Early detection and management is crucial for the control of DR and annual fundus examination (DR screening) for all diabetic patients is recommended in many developed countries. However, these are not available in China yet, because over 46% diabetic patients are living in rural area where there is limited access to these expensive fundus cameras. The emerging hand-held fundus camera, as a new point-of-care (POC) technique holds great potential to address the above problem due to its low-cost and easy-to-operate. However, the changes of retinal vasculature is very subtle, usually in a few microns and a few pixels in fundus photographs. Traditional diagnose method depends on the experience of clinicians and their trained image reading skills, which is time-consuming, expensive, and prone to error. So computer-aided diagnosis (CAD) algorithms have been widely needed in the quantification, analysis, and understanding of fundus photographs. However, the new fundus photograph is very different from the traditional fundus photograph. It has a narrower field-of-view (FOV) and lower image contrast, meaning the existing algorithms cannot be applied to the new fundus photograph directly. Here we propose to develop a fully automatic retinal vasculature analysis system for the new fundus photograph and use this system to further study the relationship between the changes of retinal vasculature and the development of DR. State-of-the-art medical image analysis algorithms such as machine learning, graph theory, and SIFT methods are used to develop this system. The goal is to provide the technique and standard for computer-aided DR diagnosis in rural China. It is also important in the sense of providing support for wide range DR screening in the future.

糖尿病视网膜病变（diabetic retinopathy）是糖尿病的严重并发症之一。眼底图像检查是视网膜病变早期诊断和管理的主要方法，传统的临床诊断依赖于医生的经验和读片能力，但眼底图像上视网膜血管变化非常细微，传统诊断方法耗时耗力且主观性强，因此亟需一种计算机辅助诊断方法。本项目拟针对一种新型眼底图像，建立系统的视网膜血管定量分析方法，采用多维图像纹理特征及色彩特征改善低对比度情况下的动脉／静脉分类，并采用图论方法对视网膜血管进行亚像素级的数学建模，使视网膜血管的测量精度大幅提高；通过血管参数与视网膜病变其他成熟诊断参数相关性的深入研究，进一步阐明视网膜血管径改变与糖尿病视网膜病变的关联，为糖尿病视网膜病变的计算机辅助诊断提供理论依据和技术支持。本项目为普及糖尿病视网膜病变在我国的早期诊断和大规模筛查提供了新的研究思路。

糖尿病视网膜病变是糖尿病的严重并发症之一，是成年人的主要致盲性眼病。对糖尿病患者的早期诊断和管理是降低致盲的关键环节。眼底图像检查是糖尿病视网膜病变早期诊断和管理的主要方法，其中血管管径的改变是糖尿病视网膜病变的一个重要特征。然而视网膜血管管径变化细微，传统基于医生读片的诊断方法耗时耗力且主观性强，因此亟需一种计算机辅助诊断的方法。本项目的研究一种新型眼底图像，对系统实现血管定量分析所涉及的三个核心算法进行研究，建立了系统的视网膜血管定量分析方法，并以此为基础深入探索了视网膜血管的改变与糖尿病视网膜病变及其他糖尿病微血管并发症的关系。围绕研究目标，本项目一共展开了一下六项工作：.1） 建立临床数据库2个：与西安市第一医院合作，采集正常组眼底图像和便携眼底照相机图像各30例，用于针对便携眼底照相机的算法开发。与西京医院内分泌科展开合作，共采集临床横断面糖尿病人数据2066例，采集的临床数据包括左右眼眼底彩照各一张及各项生理生化检查指标50余项，并对血管参数与包括糖尿病视网膜病变在内的多种微血管并发症的相关性进行了研究。.2） 高精度、低运算要求的血管分割算法研究，建立基于显著性特征的血管分割方法，该方法运算复杂度低，鲁棒性高，精确度高。.3） 基于传统机器学习方法的视网膜血管动脉/静脉分类算法研究，建立基于多种图像特征的血管分割方法，创新性地提出通过图像内和图像间均一化，减少了特征空间的差异，提高了方法的准确率。.4） 基于卷积神经网络的动脉/静脉分割算法研，建立基于深度学习的高精度动脉/静脉分割方法，大幅度提高了算法的精度。.5）利用全自动血管分析方法，对视网膜血管定量参数与糖尿病微血管病变的相关性进行探索：与西京医院合作，利用临床糖尿病人数据库中的数据，对血管参数与糖尿病微血管并发症的相关性进行研究。基于目前的数据，首先对血管参数和糖尿病肾病的相关性进行了探索。研究结果显示，血管分型维数的降低更容易出现糖尿病肾病，且随着FD的减小，患病风险依次增加。同样的，血流灌注参数与eGFR之间均有显著的负相关性。也就是说无血管灌注的区域越大，eGFR越小，更容易出现糖尿病肾病。.6） 设计并实现了多款针对便携眼底照相机图像的视网膜血管自动分析手机App以及多种用于临床研究的基于Windows的软件。