基于深度卷积神经网络的早产儿视网膜病变1~3期分期自动诊断

Automatic Diagnosis of Stages 1-3 Retinopathy of Prematurity Based on Deep Convolutional Neural Network

扫码查看

原文链接

万方数据

中文摘要：目的：利用深度卷积神经网络（DCNN）对早产儿视网膜病变诊断分期中1~3期病变进行自动分类诊断。方法：回顾性研究。选取2019年1月至2020年12月在嘉兴市妇幼保健院出生的1 885例早产儿所采集的12 219张眼底图像，构建了早产儿视网膜眼底图像数据集。基于分割出眼底图像的视网膜血管以及分界线或嵴，计算出感兴趣区域（ROI），并从ROI分割图中提取特征，采用五折交叉验证法训练分类器，对早产儿视网膜病变中1~3期进行自动分类诊断。在测试集上对DCNN进行性能评估，并与临床诊断结果进行一致性分析。结果：本系统对ROP 1~3期及无ROP的诊断准确率达到了98%。在诊断无ROP图像时，其敏感度和特异性分别达到了0.975 7和0.975 6；对ROP 1期、2期和3期图像的诊断，敏感度分别为0.922 1、0.933 1和0.910 2，特异性则分别为0.983 7、0.988 6和0.992 8。DCNN的诊断结果与临床诊断结果的Kappa一致性为0.905 9。结论：基于DCNN的早产儿眼底病变分期诊断系统，使用从ROI分割图中提取的特征训练分类器，能够对ROP1~3期病变眼底图像进行较高准确率的自动辅助诊断。 Objective: The aim of the system is to research the automatic diagnosis of stages 1-3 of retinopathy of prematurity(ROP) using Deep Convolutional Neural Networks(DCNN). Methods: In this retrospective study, using 12 219 retinal images of preterm infants, which collected from January 2019 to December 2020 at the Department of Ophthalmology, Jiaxing Maternity and Child Health Care Hospital, we constructed a retinal images dataset for Ophthalmology of Ophthalmology, preterm infants. Based on the segmented demarcation lines or ridge, the region of interest (ROI) were calculated, features from the ROI segmentated images were extracted and the classifier was trained using a five-fold cross-validation method to automatically diagnose stages 1-3 ROP. The performance of the DCNN and analyzed the consistency with clinical diagnosis results on the test data set was evaluated. Results: The trained system achieved an average accuracy of 98% for all the four categories. The sensitivity and specificity of the system reached 0.975 7 and 0.975 6, when diagnosing non-ROP images 0.922 1 and 0.983 7, when diagnosing stage 1 0.933 1 and 0.988 6, when diagnosing stage 2. At the same time, the sensitivity and specificity for the diagnosis of stage 3 ROP images were as high as 0.910 2 and 0.992 8. The Kappa value of the system for the diagnosis was 0.905 9, which was close to perfect agreement with the clinic diagnosis. Conclusion: The system based on DCNN, trained using features extracted for segmented ROI images, could diagnose automatically stages 1-3 ROP with a high accuracy.

外文摘要：Objective: The aim of the system is to research the automatic diagnosis of stages 1-3 of retinopathy of prematurity(ROP) using Deep Convolutional Neural Networks(DCNN). Methods: In this retrospective study, using 12 219 retinal images of preterm infants, which collected from January 2019 to December 2020 at the Department of Ophthalmology, Jiaxing Maternity and Child Health Care Hospital, we constructed a retinal images dataset for Ophthalmology of Ophthalmology, preterm infants. Based on the segmented demarcation lines or ridge, the region of interest (ROI) were calculated, features from the ROI segmentated images were extracted and the classifier was trained using a five-fold cross-validation method to automatically diagnose stages 1-3 ROP. The performance of the DCNN and analyzed the consistency with clinical diagnosis results on the test data set was evaluated. Results: The trained system achieved an average accuracy of 98% for all the four categories. The sensitivity and specificity of the system reached 0.975 7 and 0.975 6, when diagnosing non-ROP images 0.922 1 and 0.983 7, when diagnosing stage 1 0.933 1 and 0.988 6, when diagnosing stage 2. At the same time, the sensitivity and specificity for the diagnosis of stage 3 ROP images were as high as 0.910 2 and 0.992 8. The Kappa value of the system for the diagnosis was 0.905 9, which was close to perfect agreement with the clinic diagnosis. Conclusion: The system based on DCNN, trained using features extracted for segmented ROI images, could diagnose automatically stages 1-3 ROP with a high accuracy.

外文关键词：

retinadeep convolutional neural networkspremature infantimages segmentation

作者：

刘佳、濮清岚、李鹏、周巧云、许维馨、李勇、吴昔昔

展开 >

作者单位：

1嘉兴市妇幼保健院眼科，嘉兴　314000

2同济大学浙江学院电子与信息工程系，嘉兴　314005

关键词：

视网膜病变深度卷积神经网络早产儿图像分割

基金：

浙江省医药卫生科技面上项目嘉兴市科技计划项目Zhejiang Province Medical and Health Science and Technology General ProjectJiaxing Science and Technology Project

项目编号：

2020KY9652020AD300412020KY9652020AD30041

出版年：

2022

DOI：

10.3760/cma.j.cn115909-20220307-00087

中华眼视光学与视觉科学杂志

中华医学会

中华眼视光学与视觉科学杂志

CSTPCDCSCD

影响因子：0.783

ISSN：1674-845X

年,卷(期)：2022.24(12)

参考文献量2