上海交通大学学报（医学版）

上海交通大学学报（医学版） ›› 2019, Vol. 39 ›› Issue (6): 598-.doi: 10.3969/j.issn.1674-8115.2019.06.007

• 论著·临床研究 • 上一篇    下一篇

标准七视野眼底彩色照相中微动脉瘤病灶自动化识别研究

余奇 1, 2，刘梦雪 3，杨杰 3，刘堃 1, 2，许迅 1, 2   

  1. 1. 上海交通大学附属第一人民医院眼科，上海 200080；2. 上海交通大学眼科研究所读片中心，上海 200080；3. 上海交通大学图像分析及模式识别研究所，上海 200240
  • 出版日期:2019-06-28 发布日期:2019-07-26
  • 通讯作者: 许迅，电子信箱：drxuxun@sjtu.edu.cn。
  • 作者简介:余奇（1987—），女，主治医师，博士；电子信箱： yu.qi@sjtu.edu.cn。
  • 基金资助:
    国家重点研发计划资助项目（2016YFC0904800）；国家自然科学基金（81600776）；上海市科学技术委员会医学引导类项目（16411962100）

Automated analysis of microaneurysm lesions in standard seven-field color fundus photography

YU Qi1, 2, LIU Meng-xue3, YANG Jie3, LIU Kun1, 2, XU Xun1, 2   

  1. 1. Department of Ophthalmology, Shanghai General Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai 200080, China; 2. Eye Institute Reading Center, Shanghai Jiao Tong University, Shanghai 200080, China; 3. Institute of Image Processing and Pattern Recognition, Shanghai Jiao Tong University, Shanghai 200240, China
  • Online:2019-06-28 Published:2019-07-26
  • Supported by:
    National Key R&D Plan of China, 2016YFC0904800; National Natural Science Foundation of China, 81600776; Guiding Medical Project of Shanghai Science and Technology Committee, 16411962100）。

PDF (PC)

398

摘要： 目的 ·评估监督机器学习算法自动化检测标准七视野眼底彩色照相中微动脉瘤病灶的准确性及检测效率。方法 ·采集录入 2014—2016年于上海交通大学附属第一人民医院眼科门诊就诊的 44例糖尿病性视网膜病变（diabetic retinopathy，DR）患者的标准七视野眼底彩色照相图像 616幅。通过组合应用包括图像预处理、双窗口滤波器、支持向量机等 5个关键步骤，对微动脉瘤病灶实现自动化检测，同时与眼科专科医师的手工标记进行比对，评估自动化算法对微动脉瘤病灶识别的准确性及检测效率。结果 ·在 DR标准七视野眼底彩色照相图像库中，计算机自动化识别算法的微动脉瘤检测灵敏度为 94.15%、特异度为 98.05%。其中，在视盘视野、黄斑视野、黄斑颞侧视野、颞上视野、颞下视野、鼻上视野、鼻下视野各图像分集合中的算法检测灵敏度分别为 93.09%、94.84%、 95.16%、94.99%、93.77%、92.40%、93.75%，特异度分别为 98.02%、98.06%、97.97%、97.91%、98.07%、98.03%、98.23%，算法在各图像分集合中的微动脉瘤病灶检测灵敏度和特异度同总集合中的检测灵敏度差异没有统计学意义（P>0.05）。每幅图像检测耗时（9.2±0.6）s，较手工标记节约用时 93.2%。结论 ·基于监督机器学习算法的微动脉瘤病灶自动识别算法能够准确、高效地识别标准七视野眼底彩色照相中及各个视野范围中的微动脉瘤病灶。

关键词: 糖尿病性视网膜病变, 微动脉瘤, 模式识别, 图像分析, 眼底彩色照相

Abstract:

Objective · To evaluate the accuracy and efficiency of the automated supervised machine-learning algorithm for microaneurysm lesion detection in seven-field color fundus photography. Methods · A total of 616 seven-field color fundus photographs were obtained 44 patients with diabetic retinopathy (DR) Department of Ophthalmology, Shanghai General Hospital, Shanghai Jiao Tong University School of Medicine 2014 to 2016. Using the microaneurysm detection algorithm developed in this study, the automated identification and labeling of microaneurysm lesions in the standard seven-field color photography of DR were performed. The results were compared with manual labelingophthalmologists to evaluate the sensitivity and efficiency of the automated algorithm. Results · In the standard seven-field fundus color photographic image library, the automated algorithm achieved sensitivity of 94.15% in total and 93.09% in the optic disc field (F1), 94.84% in the macula field (F2), 95.16% in the temporal to macula field (F3), 94.99% in the superior temporal field (F4), 93.77% in the inferior temporal field (F5), 92.40% in the superior nasal field (F6) and 93.75% in the inferior nasal field (F7), and specificity of 98.05% in total and 98.02% in F1, 98.06% in F2, 97.97% in F3, 97.91% in F4, 98.07% in F5, 98.03% in F6 and 98.23% in F7. The cost of time per image was (9.2± 0.6) s, 93.2% less time than manual labeling. Conclusion · The automated microaneurysm detection algorithm can accurately and efficiently identify microaneurysm lesions in color fundus photography.

Key words: diabetic retinopathy (DR), microaneurysm, pattern recognition, image analysis, colour fundus photography

中图分类号: