基于眼底图像非结构化数据建模的糖网病自动筛查研究

Diabetic retinopathy is the most serious complications of diabetic microangiopathy, and it has become the primary blind factor. To seek effective computer-aided screening solutions of diabetic retinopathy are imperative. Computer-aided screening can greatly reduce the risk of blindness in diabetic patients, however, out-of-state methods of screening are difficult to clinical practical for high error of mistake detection.This project proposed a high-efficiency computer-aided screening programm which is based on medical image processing. First of all, we create a new Chinese retinal fundus image database, which include health and diabetic, construct training sample set of retinal vessels and lesions. Then the retinal blood vessels and vascular abnormalities and lesions in fundus images are processed based on digital image processing and supervised learning. Features of retinal blood vessels and lesions are extracted which consist of gray and edge. Unstructured data modeling of retinal vessels, automatic detection and quantization of lesions are important processes in this research. Finally with the multiple decision fusion device, we can get conclusion whether there is pathology in the retinal. Project aims to provide a solution of diabetic retinopathy computer-aided screening, for more than million people with diabetes in China, to prevent of blindness caused by diabetes, to alleviate the current situation of the shortage of medical resources.

糖网病是糖尿病性微血管病变最严重的并发症，已成为首要致盲因素，寻求高效的糖网病自动筛查解决方案已刻不容缓。计算机辅助的眼底图像筛查可以大幅度降低糖尿病患者的失明风险，然而已有的自动筛查方法因误筛、漏筛率偏高，难以临床实用而备受关注。本项目拟基于医学图像分析处理技术，提出一种快速精准的糖网病自动筛查方案。首先针对中国人进行健康人和糖尿病患者的眼底图像采集建库工作，构建眼底图像训练样本集；然后以眼底图为研究对象，以数字图像处理技术、监督学习为核心方法，深入分析视网膜血管、血管异常以及病灶的形态、灰度、边缘等特征，对视网膜血管等非结构化数据建模，自动检测并量化病灶；最后采用多决策器融合的方式，得到待查视网膜是否病变的精确判断，实现糖网病的自动筛查。项目旨在为我国逾亿糖尿病患者提供定期的糖网病全覆盖筛查服务，有效防治因糖网病导致的失明，缓解医疗资源紧缺的现状。

糖网病是糖尿病最常见的视网膜微血管病变，是首要致盲因素，其早期的筛查和治疗至关重要。但目前糖网病的诊断主要依赖于眼科医生的肉眼阅片，工作量大，效率低。对于难以发现的微小病灶存在一定的漏检。已有的自动筛查方法因误筛、漏筛率偏高，难以临床实用而备受关注。为此，本项目利用医学影像分析技术，结合机器学习，提出一种快速精准的糖网病自动筛查方案，对眼底的非结构化数据进行建模分析。项目的研究内容包括：建立针对中国人的彩色眼底图数据库；视网膜血管分割与建模；糖网病视网膜病灶检测与量化以及糖网病自动筛查系统及临床验证四个方面。. 项目完成了眼底影像数据的采集，通过质量筛选，构建了相关的数据库，并开发了具有仿射不变性的视网膜图像配准方法。项目实现了眼底血管，视杯与视盘的结构化建模，开发了能有效改善眼底细小血管提取精度的血管分割算法和基于验证模型的视盘视杯定位与分割算法；通过特征提取、选择与机器学习技术，开发了有效的动静脉分类算法，微动脉瘤检测算法，糖网筛查算法；最终构建了一个糖网病交互筛查平台，用于各类算法的验证。项目组共发表论文11篇，申请专利6项，获得软件著作权1项。. 针对眼底影像的关键生理结构进行建模和分析，挖掘其中的有效信息，结合糖网病病因进行分析，能更加客观，科学地得出糖网病筛查结果，给临床医生参考，为广大眼疾患者带来福音，助推健康中国。