摘要
目的研究基于卷积神经网络自动检测青光眼性视神经病变的深度学习算法,并探讨实现病灶区可视化的可行性。设计横断面研究。研究对象2014-2018年北京同仁医院5148例患者10296眼的眼底图像。方法在提供有无青光眼性视神经病变作为标记的前提下,首先基于ResNet深度模型训练一个深度神经网络,使用训练好的模型测试并计算其诊断分类的准确性。其次利用t-分布随机邻域嵌入可视化方法(t-SNE)实现对不同类别的深度特征分布可视化,生成相应的病灶区域热力图。计算该深度学习算法分类的敏感性、特异性和受试工作特性曲线下面积(AUC),并通过热力图评价其对某种类型病灶区的识别准确率以及对于诊断贡献最大的区域与专家的判别一致性。主要指标敏感性、特异性、AUC、识别准确率、判别一致性。结果在验证数据集中,该算法的AUC为0.996(95%CI,0.995-0.998),检测到病灶区的敏感性和特异性与受过培训的专业评分员相当(敏感性,96.2%vs 96.0%,P=0.76;特异性,97.7%vs 97.9%,P=0.81)。病灶区域热力图对视盘异常和盘沿丢失区域的识别准确率达到100%,对于诊断贡献最大的区域判别与青光眼专家的一致性达91.8%。结论运用深度学习算法检测青光眼性视神经病变的眼底图像具有较高的敏感性与特异性,同时基于t-SNE算法实现了对诊断贡献较大的病灶区域可视化。
Objective To study a deep learning system(DLS)based on convolutional neural network(CNN)for automated detection of glaucomatous optic neuropathy,and to perform a prediction visualization test that can identify regions of the fundus images.Design Cross-sectional study.Participants Ocular fundus photos of 10296 eyes of 5148 patients during 2014 to 2018 in Beijing Tongren Hospital.Methods A deep learning algorithm based on ResNet was trained on the premise that only disease or not can be provided as a marker,then the accuracy,sensitivity and specificity of the algorithm were calculated to evaluate the performance of the trained system for automatic diagnosis.To better understand the process,a prediction visualization test was performed based on a t-distributed stochastic neighbor embedding(t-SNE)visualization that identified regions of the fundus images utilized for diagnosis,and a heatmap was created.Main Outcome Measures Area under the receiver operating characteristics curve(AUC),sensitivity and specificity for DLS with reference to professional graders,diagnosis accuracy and consistency with ophthalmologists according to the heatmap.Results The AUC of the glaucoma diagnosis with CNN(GD-CNN)model in validation datasets was 0.996(95%CI,0.995-0.998).The sensitivity and specificity were comparable with that of trained professional graders(sensitivity,96.2%vs.96.0%,P=0.76;specificity,97.7%vs.97.9%,P=0.81).An accuracy of 100%was presented in areas containing optic nerve head variance and neuroretinal rim loss,and the regions of interest identified to have made the greatest contribution to the neural network’s diagnosis were also shared with 91.8%of ophthalmologists.Conclusion The DLS has high sensitivity and specificity for detecting glaucomatous optic neuropathy.Based on t-SNE,visualization maps are generated from deep features,which can be superimposed on the input image to highlight the areas of the model important for diagnosis.
作者
张悦
庞睿奇
杜一帆
牟大鹏
李柳
徐迈
王宁利
刘含若
Zhang Yue;Pang Ruiqi;Du Yifan;Mu Dapeng;Li Liu;Xu Mai;Wang Ningli;Liu Hanruo(Beijing Institute of Ophthalmology,Beijing Tongren Eye Center,Beijing Key Laboratory of Ophthalmology and Visual Sciences,Beijing Tongren Hospital,Capital Medical University,Beijing 100005,China;Beihang University,Beijing 100191,China)
出处
《眼科》
CAS
2020年第1期9-14,共6页
Ophthalmology in China
基金
国家自然科学基金(81700813)
北京市医院管理局“青苗”计划专项经费(QML20180205)
首都医科大学附属北京同仁医院种子基金项目(2016-YJJ-ZZL-021)
首都医科大学附属北京同仁医院拔尖人才培养计划,医药协同科研创新研究专项(Z181100001918035)。
关键词
人工智能
卷积神经网络
青光眼性视神经病变
artificial intelligence
convolutional neural network
glaucomatous optic neuropathy
