北京生物医学工程

基于深度迁移学习的老年人群眼底疾病辅助诊断研究

Fundus diseases of the elderly detection algorithm based on deep transfer learning

作者：杨康赵太宏

单位：南京医科大学公共卫生学院（南京 211112）;<br />南京医科大学附属南京医院院长办公室（南京 210006）<br />通信作者：赵太宏。E-mail:njchbme@163.com

分类号： R318.04 

出版年·卷·期（页码）：2022·41·5（465-470）

摘要：

目的提出一种基于深度迁移学习的老年人群眼底疾病自动分类方法，为老年人眼底疾病提供辅助诊断分析。方法以2 048张眼底图像作为原始眼底数据集。首先，对眼底图像进行去除模糊不清、图像缩放和对比度调整等预处理操作。然后随机将眼底图像划分成训练组、验证组和测试组。迁移卷积神经网络（convolution neural network，CNN）在其他分类任务预训练的参数，然后根据眼底病分类任务微调CNN模型全连接层参数，经过网络优化最终搭建眼底病分类诊断网络模型。最后使用独立的测试集对网络模型进行测试，并通过混淆矩阵查看模型的分类情况。最终，以受试者工作特征曲线下面积（area under curve，AUC）、准确率、灵敏度、特异度和F1值等指标评价网络模型对眼底病的分类诊断效果。结果 CNN模型在眼底病分类准确率均高于0.90，灵敏度在0.90~0.95范围，特异度在0.89~0.95范围，AUC在0.87~0.92范围，F1值在0.92~0.94范围。结论提出基于迁移学习方法训练的CNN模型以较高的准确率实现老年人眼底病的自动分类诊断，且具备训练周期短和训练参数少等优势。该方法将有助于提高基层社区对老年人群眼底疾病的辅助诊断能力。

Objective To propose an automatic classification method of fundus diseases for the elderly based on deep transfer learning, and to provide auxiliary diagnosis and analysis of fundus diseases for the elderly. Methods Total of 2048 fundus images were used as the original fundus dataset. First, the fundus image is preprocessed, such as removing blur, image scaling and contrast adjustment.Then,The fundus images were randomly divided into training group, validation group and testing group. The parameters of the convolutional neural network (CNN) pre-trained in other classification tasks were transferred, and then the parameters of the fully connected layer of the CNN model were fine-tuned according to the fundus disease classification task. After network optimization, the fundus disease classification and diagnosis network model was finally built. Finally, an independent test set was used to test the network model, and the classification of the model was checked through the confusion matrix. At last, the classification and diagnosis effect of the network model on fundus diseases was evaluated by indicators such as area under curve (AUC), accuracy, sensitivity, specificity and F1 score. Results The classification accuracy of CNN model in fundus diseases was higher than 0.90, the sensitivity was in the range of 0.90-0.95, the specificity was in the range of 0.89-0.95, the AUC was in the range of 0.87-0.92, and the F1 score was in the range of 0.92-0.94. Conclusions The CNN model trained based on the transfer learning method is proposed to achieve automatic classification and diagnosis of fundus diseases in the elderly with high accuracy, and has the advantages of short training period and few training parameters. This method will help to improve the auxiliary diagnosis ability of the elderly population for fundus diseases in the grass-roots community.

参考文献：

[1] 高华, 陈秀念, 史伟云. 我国盲的患病率及主要致盲性疾病状况分析[J]. 中华眼科杂志, 2019, 55(8): 625-628.
Gao H,Chen XN,Shi WY. Analysis of the prevalence of blindness and major blinding diseases in China[J].Chinese Journal of Ophthalmology,2019,55(8):625-628
[2] 佟甜, 姜艳华. 老年性黄斑变性发病率及危险因素分析[J]. 国际医药卫生导报, 2019, 25(1): 14-16.
[3] 中华医学会眼科学分会眼视光学组. 重视高度近视防控的专家共识(2017)[J]. 中华眼视光学与视觉科学杂志, 2017, 19(7): 385-389
[4] 沈亚琴, 杨梅, 刘必红, 等. 江苏省糖尿病眼病研究中阜宁县50岁以上2型糖尿病患者盲和中重度视力损伤的流行病学调查[J]. 中华眼科杂志, 2020, 56(8): 593-599
Shen YQ,Yang M,Liu BH ,et al. Jiangsu diabetic eye disease study：epidemiological survey of blindness and moderate or severe visual impairment in people with type 2 diabetes over 50 years old in Funing County [J].Chinese Journal of Ophthalmology,2020,56(8):593-599.
[5] 陈战巧, 俞颂平. 浙江省南部地区畲族老年人群眼病流行病学调查研究[J]. 中国预防医学杂志, 2020, 21(10): 1099-1103.
Chen ZQ,Yu SP. The research on epidemiological investigation and preventive measures of eye diseases in the elderly of the she nationality in southern zhejiang.[J]. Chinese Preventive Medicine,2020,21(10):1099-1103.
[6] 魏串串, 刘雪, 王爽, 等. 40岁以上中老年人视网膜血管弯曲度的横断面调查——北京眼病研究[J]. 眼科, 2021, 30(2): 97-101.
Wei CC,Liu X,Wang S,et al. Cross-sectional study of retinal vascular tortuosity in elderly population-Beijing eye study.[J]. Ophthalmology in China,2021,30(2):97-101.
[7] 吴晓兰, 易全勇, 邬一楠, 等. 宁波地区50岁及以上人群眼病流行病学调查[J]. 中华全科医学, 2019, 17(3): 491-495.
[8] Xu T, Wang B, Liu H, et al. Prevalence and causes of vision loss in China from 1990 to 2019: findings from the Global Burden of Disease Study 2019[J]. The Lancet Public Health, 2020,5(12):e682-e691.
[9] Dewey M, Schlattmann P. Deep learning and medical diagnosis[J]. Lancet, 2019,394(10210):1710-1711.
[10] Cen LP, Ji J, Lin JW, et al. Automatic detection of 39 fundus diseases and conditions in retinal photographs using deep neural networks[J]. Nature Communications, 2021,12(1):4828.
[11] Son J, Shin JY, Kim HD, et al. Development and validation of deep learning models for screening multiple abnormal findings in retinal fundus images[J]. Ophthalmology, 2020,127(1):85-94.
[12] 高爽, 徐巧枝. 迁移学习方法在医学图像领域的应用综述[J]. 计算机工程与应用, 2021,57(24): 39-50.
Gao S, Xu QZ.Review of application of transfer learning inmedical image field[J].Computer Engineering and Applications,2021, 57(24): 39-50
[13] 赵蒙蒙, 鲁贞贞, 朱书缘, 等. 基于卷积神经网络的眼科光学相干断层成像图像的自动分类[J]. 北京生物医学工程, 2021, 40(6): 557-563.
Zhao MM, Lu ZZ, Zhu SY,et al. Automatic classification of ophthalmic optical coherence tomography images based on the convolution neural network[J].Beijing Biomedical Engineering, 2021, 40(4):557-563.
[14] Zhang K, Sun M, Han TX, et al. Residual networks of residual networks: multilevel residual networks[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2018,28(6):1303-1314.
[15] Ran W, Fu K, Hao S, et al. Image superresolution using densely connected residual networks[J]. IEEE Signal Processing Letters, 2018,25(10):1565-1569.
[16] Kanavati F, Toyokawa G, Momosaki S, et al. Weakly-supervised learning for lung carcinoma classification using deep learning[J]. Scientific Reports, 2020,10(1):9297.
[17] Kermany DS, Goldbaum M, Cai W, et al. Identifying medical diagnoses and treatable diseases by image-based deep learning[J]. Cell, 2018,172(5):1122-1131.
[18] Gargeya R, Leng T. Automated identification of diabetic retinopathy using deep learning[J]. Ophthalmology, 2017,124(7):962-969.
[19] Gulshan V, Peng L, Coram M, et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs[J]. JAMA, 2016,316(22):2402-2410.
[20] Hua CH, Kim K, Huynh-The T, et al. Convolutional network with twofold feature augmentation for diabetic retinopathy recognition from multi-modal images[J]. IEEE Journal of Biomedical and Health Informatics, 2021,25(7):2686-2697.

服务与反馈：

【文章下载】【加入收藏】

提示：您还未登录，请登录！点此登录