电子学报 ›› 2019, Vol. 47 ›› Issue (12): 2611-2621.DOI: 10.3969/j.issn.0372-2112.2019.12.022

• 学术论文 • 上一篇    下一篇

基于多模型融合和区域迭代生长的视网膜血管自动分割

赖小波1, 许茂盛2, 徐小媚1   

  1. 1. 浙江中医药大学医学技术学院, 浙江杭州 310053;
    2. 浙江中医药大学第一临床医学院, 浙江杭州 310006
  • 收稿日期:2018-05-18 修回日期:2019-06-10 出版日期:2019-12-25
    • 作者简介:
    • 赖小波 男,1981年生于江西赣州,博士,浙江中医药大学医学技术学院副教授,主要研究方向为数字医学影像处理与分析.E-mail:dmia_lab@zcmu.edu.cn;许茂盛 男,1966年生于浙江杭州,博士,浙江中医药大学第一临床医学院主任医师,主要研究方向为医学影像诊断学.E-mail:xms@sina.com
    • 基金资助:
    • 国家自然科学基金 (No.61602419); 浙江省自然科学基金 (No.LY16F10008,No.LQ16F020003)

Automatic Retinal Vessel Segmentation Based on Multi-model Fusion and Region Iterative Growth

LAI Xiao-bo1, XU Mao-sheng2, XU Xiao-mei1   

  1. 1. Medical Technology College, Zhejiang Chinese Medical University, Hangzhou, Zhejiang 310053, China;
    2. First Clinical Medicine College, Zhejiang Chinese Medical University, Hangzhou, Zhejiang 310006, China
  • Received:2018-05-18 Revised:2019-06-10 Online:2019-12-25 Published:2019-12-25
    • Supported by:
    • National Natural Science Foundation of China (No.61602419); National Natural Science Foundation of Zhejiang Province,  China (No.LY16F10008, No.LQ16F020003)

摘要: 糖尿病视网膜病变是成年人致盲首因,视网膜血管分割是诊断糖尿病视网膜病变的基础.为提高视网膜血管分割准确性,提出一种基于多模型融合和区域迭代生长的视网膜血管自动分割算法.首先,预处理后分别构建数学形态学、匹配滤波器、尺度空间分析、多尺度线检测和神经网络模型初步分割视网膜血管,为减少噪声取五个分割结果的均值作为初步输出.其次,设计掩膜分离渗出物和视盘,将数学形态学模型分割结果替换掩膜白色区域,并融合初步输出生成组合结果.最后,考虑视网膜血管先验知识,对组合结果阈值分割和区域迭代生长后获取最终结果.实验结果表明,该算法分割DRIVE和STARE眼底图像库视网膜血管的检测精度、敏感度和特异性分别为0.9457、0.7843、0.9815以及0.9472、0.7826、0.9803,优于多数经典算法.

关键词: 视网膜血管, 自动分割, 多模型融合, 区域迭代生长

Abstract: Diabetic retinopathy is the leading cause of blindness in adults, and the retinal vessels segmentation is the basis for the diagnosis of diabetic retinopathy. To improve the accuracy of retinal vessels segmentation, an automatic retinal vessel segmentation method based on multi-model fusion and region iterative growth is proposed. Firstly, the mathematical morphology, matched filter, scale-space analysis, multi-scale line-detection and neural network models were established to segment retinal vessels initially, and the mean value of the five segmentation results was taken as the preliminary output to reduce the noise. Secondly, after exudates and optic disc were separated with a mask developed, the white areas in the mask were replaced by the segmentation result of the mathematical morphology model, and then the combined result was generated with replaced mask and preliminary output fused. Finally, considering the prior knowledge of retinal vessels, final results were obtained after threshold and region iterative growth. The experimental results demonstrate that the accuracy, sensitivity and specificity for segmenting the retinal vessels in the DRIVE and STARE fundus image datasets are 0.9457, 0.7843, 0.9815 and 0.9472, 0.7826 and 0.9803, respectively, which is superior to most classical algorithms.

Key words: retinal vessel, automatic segmentation, multi-model fusion, region iterative growth

中图分类号: