切换至 "中华医学电子期刊资源库"
高级检索
中华眼科医学杂志(电子版)

中华眼科医学杂志(电子版) ›› 2020, Vol. 10 ›› Issue (04) : 234 -238. doi: 10.3877/cma.j.issn.2095-2007.2020.04.007

论著

人工智能深度学习技术在辅助青光眼性眼底病变图像标注中的应用研究
刘含若1, 白玮玲1, 张悦1, 杜一帆1, 王宁利1,†()   
  1. 1. 100730，首都医科大学附属北京同仁医院　北京同仁眼科中心　北京市眼科研究所　北京市眼科学与视觉科学重点实验室
  • 收稿日期:2020-08-11 出版日期:2020-08-28
  • 通信作者: 王宁利
  • 基金资助:
    国家自然科学基金项目(81700813); 北京市医院管理局"青苗"计划专项经费项目(QML20180205); 北京市科技新星项目(Z191100001119072); 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)

Application of artificial intelligence deep learning technology to assist annotation of glaucomatous fundus

Hanruo Liu1, Weiling Bai1, Yue Zhang1, Yifan Du1, Ningli Wang1,†()   

  1. 1. Beijing Tongren Eye Cener, Beijing Tongren Hospital, Capital Medical University, Beijing Institute of Ophthalmology, Beijing Ophthalmology & Visual Sciences Key Lab., Beijing 100730, China
  • Received:2020-08-11 Published:2020-08-28
  • Corresponding author: Ningli Wang
全文 ( ) 下载PDF ( ) 导出引用
引用本文：

刘含若, 白玮玲, 张悦, 杜一帆, 王宁利. 人工智能深度学习技术在辅助青光眼性眼底病变图像标注中的应用研究[J]. 中华眼科医学杂志(电子版), 2020, 10(04): 234-238.

Hanruo Liu, Weiling Bai, Yue Zhang, Yifan Du, Ningli Wang. Application of artificial intelligence deep learning technology to assist annotation of glaucomatous fundus[J]. Chinese Journal of Ophthalmologic Medicine(Electronic Edition), 2020, 10(04): 234-238.

目的

探索人工智能(AI)深度学习技术用于青光眼性眼底病变(GON)筛查检出的效能。

方法

采用诊断试验的研究方法，收集自2013年8月至2019年7月于首都医科大学附属北京同仁医院北京同仁眼科中心青光眼门诊就诊患者的眼底图像200张(200只眼)。根据阅片方式的不同，将采用AI系统阅片的定义为AI系统组，将医师人工阅片的定义为医师组。其中，医师组包括高年资眼科医师(眼底照片诊断经验为10年以上)、低年资眼科医师(眼底照片诊断经验为5年以上，10年以内)及全科医师。各组均对所有眼底图像进行阅片和标注。年龄和性别的数据经S-W检验证实呈正态分布者，以均数±标准差表示。经Levene检验证实方差齐，AI系统与不同级别医师单张平均阅片时间的比较采用单因素方差分析，两两比较采用LSD-t检验。采用敏感度、特异度及工作特性曲线下面积(AUC)对AI模型进行预测和性能评价。诊断符合率以百分数表示，AI系统与眼科高级医师阅片诊断的符合率、AI系统与不同级别医师阅片诊断符合率的比较，均采用卡方检验。

结果

在医师组中，高年资眼科医师、低年资眼科医师及全科医师共6位医师共完成2次阅片。其诊断符合率的比较，差异均有统计学意义(χ2＝4.324，3.562，4.213，5.786，10.546，11.431；P<0.05)。AI系统阅片测试的敏感度为100.0%，特异度为88.6%，诊断符合率为90.5%，AUC为0.934。在AI系统辅助下，标注的符合率提升最大的为2位全科医师，分别达到了87.3%及82.5%。在AI系统辅助下，高年资眼科医师、低年资眼科医师及全科医师的平均阅片时间较独立阅片标注有明显缩短，差异均有统计学意义(t＝4.175、3.189、6.160；P<0.05)。

结论

标准化智能AI标注辅助系统的应用有利于不同资质医师的准确诊断和提高效率。尤其，可以快速提升全科医师及低年资眼科医师标注的符合率。

关键词: 人工智能, 青光眼, 眼底图像
Objective

To explore the effectiveness of artificial intelligence (AI) deep learning technology in the screening and detecting glaucomatous optic neuropathy (GON).

Methods

Using diagnostic test research method, 200 fundus images (200 eyes) were collected from the glaucoma clinic of Beijing Tongren Eye Center from August 2013 to July 2019. According to the different ways of image reading, they were divided into two groups: AI system group and doctor group. Among of them, the doctor group included senior ophthalmologists (with more than 10 years′ experience in fundus image diagnosis), junior ophthalmologists (more than 5 years and less than 10 years) and general practitioners. All fundus images were read and labeled by AI system group and doctor group. The data of age and sex were normal distribution by S-W test, expressed as (±s). As the Levene test showed that the real variance was the same, the average image reading time of AI system was compared with that of doctors of different levels by one-way ANOVA, and the LSD-t test was used for pairwise comparison. Sensitivity, specificity and area under curve (AUC) were used to predict and evaluate the performance of AI model. The diagnostic coincidence rate was expressed as a percentage. Chi square test was used to compare the diagnostic coincidence rate between AI system and senior ophthalmologists, between AI system and doctors at different levels.

Results

In the doctor group, 6 ophthalmologists (senior ophthalmologists, junior ophthalmologists and general practitioners) read images twice. The difference of diagnostic accuracy was statistically significant (χ2=4.324, 3.562, 4.213, 5.786, 10.546, 11.431; P<0.05). The sensitivity, specificity, diagnostic coincidence rate and AUC of AI image reading test were 100.0%, 88.6%, 90.5% and 0.934, respectively. With the help of AI system, the accuracy of labeling was improved significantly in 2 general practitioners, which reached 87.3% and 82.5%, respectively. With the aid of AI system, the average reading time of senior ophthalmologists, junior ophthalmologists and general practitioners was significantly shorter than that of independent film reading (t=4.175, 3.189, 6.160; P<0.05).

Conclusions

The application of standardized intelligent AI labeling assistant system is helpful for improving the accurate rate of diagnosis and effectiveness of doctors with different qualifications. In particular, it can quickly improve the accuracy of labeling by general practitioners and junior ophthalmologists.

Key words: Artificial intelligence, Glaucoma, Fundus
图1 青光眼性眼底病变眼底图像标注标准
表1 不同年资医师对青光眼性眼底病变眼底图像分类标注一致率的比较[(n(%)]
标注次数 阅片数(张) 高年资眼科医师1 高年资眼科医师2 低年资眼科医师1 低年资眼科医师2 全科医师1 全科医师2
第1次标注 200 191(95. 5) 193(96.7) 175(87.3) 148(73.8) 121(60.4) 118(59.1)
第2次标注 200 189(94.5) 183(91.5) 162(81.1) 158(78.8) 139(69.6) 119(59.4)
χ2值   4.324 3.562 4.213 5.786 10.546 11.431
P值   <0.05 <0.05 <0.05 <0.05 <0.05 <0.05
表2 人工智能系统与不同级别医师阅片灵敏度、特异度及符合率的比较(%)
医师年资 阅片数(张) 敏感度(%) 特异度(%) 符合率(%)
无人工智能辅助 有人工智能辅助 无人工智能辅助 有人工智能辅助 无人工智能辅助 有人工智能辅助
高年资眼科医师1 200 83.6 100 98.7 96.6 95.9 95.9
高年资眼科医师2 200 85.9 100 96.5 97.4 94.1 97.3
低年资眼科医师1 200 69.7 96.4 86.2 84.1 84.2 92.3
低年资眼科医师2 200 66.1 92.5 78.4 80.5 76.3 91.0
全科医师1 200 47.3 89.7 75.3 78.3 65.4 87.3
全科医师2 200 51.9 86.4 60.6 70.2 59.2 82.5
医师平均 200 67.4 94.2 82.6 84.5 79.2 87.1
人工智能系统 200   100   88.6   90.5
表3 人工智能系统及不同级别医师单张平均阅片时间的比较(±s，s)
辅助标注 阅片数(张) 医师阅片时间 人工智能系统的阅片时间
高年资眼科师 低年资眼科师 全科医师 医师平均
无人工智能辅助 200 5.86±2.98 5.02±2.40 8.31±1.68 6.17±1.82 1.02±0.78
有人工智能辅助 200 3.64±0.58 3.72±0.20 6.47±1.78 4.08±0.93  
t值   1.534 1.342 5.131 1.543  
P值   <0. 05 <0. 05 <0. 05 <0. 05  
[1]
Rajkomar A, Dean J, Kohane I. Machine Learning in Medicine[J]. N Engl J Med, 2019, 380(14): 1347-1358.
[2]
曾伟良，吴淼森，孙为军，等. 自动驾驶出租车调度系统研究综述[J]. 计算机科学，2020，47(5): 181-189.
[3]
戴震军. 人工智能技术应用于自动驾驶汽车面临的挑战及发展趋势分析[J]. 无线互联科技，2020，17(6): 162-163.
[4]
郭坦，李金丽. 人工智能技术在网络安全问题中的应用探讨[J]. 信息技术与信息化，2020(5): 210-211.
[5]
Zhang Y, Wang N, Liu H. Applications of artificial intelligence in the screening of glaucoma in China[J]. J Med Syst, 2020, 44(7): 124.
[6]
Liu H, Li L, Wormstone IM, et al. Development and validation of a deep learning system to detect glaucomatous optic neuropathy using fundus photographs[J]. JAMA Ophthalmol, 2019, 137(12): 1353-1360.
[7]
Harris J. Who owns my autonomous vehicle? Ethics and responsibility in artificial and human intelligence[J]. Camb Q Healthc Ethics, 2018, 27(4): 599-609.
[8]
Sedjelmaci H, Guenab F, Senouci SM, et al. Cyber security based on artificial intelligence for cyber-physical systems[J]. IEEE Network, 2020, 34(3): 6-7.
[9]
Li T, Gao Y, Wang K, et al. Diagnostic assessment of deep learning algorithms for diabetic retinopathy screening[J]. Information Sciences, 2019, 501: 511-522.
[10]
Ruamviboonsuk P, Cheung CY, Zhang X, et al. Artificial intelligence in ophthalmology: evolutions in Asia[J]. Asia Pac J Ophthalmol (Phila), 2020, 9(2): 78-84.
[11]
Schmidt-Erfurth U, Sadeghipour A, Gerendas BS, et al. Artificial intelligence in retina[J]. Prog Retin Eye Res, 2018, 67: 1-29.
[12]
Lim G, Bellemo V, Xie Y, et al. Different fundus imaging modalities and technical factors in AI screening for diabetic retinopathy: a review[J]. Eye Vis (Lond), 2020, 7: 21.
[13]
Ting DSW, Cheung CY, Nguyen Q, et al. Deep learning in estimating prevalence and systemic risk factors for diabetic retinopathy: a multi-ethnic study[J]. NPJ Digit Med, 2019, 2: 24.
[14]
Li L, Xu M, Liu H, et al. A Large-Scale Database and a CNN model for attention-based glaucoma detection[J]. IEEE Trans Med Imaging, 2020, 39(2): 413-424.
[15]
Weinreb RN, Aung T, Medeiros FA. The pathophysiology and treatment of glaucoma: a review[J]. JAMA, 2014, 311(18): 1901-1911.
[16]
Stevens GA, White RA, Flaxman SR, et al. Global prevalence of vision impairment and blindness: magnitude and temporal trends, 1990—2010[J]. Ophthalmology, 2013, 120(12): 2377-2384.
[17]
Bourne RR, Stevens GA, White RA, et al. Causes of vision loss worldwide, 1990-2010: a systematic analysis[J]. Lancet Glob Health, 2013, 1(6): e339-e349.
[18]
Tham YC, Li X, Wong TY, et al. Global Prevalence of glaucoma and projections of glaucoma burden through 2040 a systematic review and meta-analysis[J]. Ophthalmology, 2014, 121(11): 2081-2090.
[19]
Lecun Y, Bengio Y, Hinton G. Deep learning[J]. Nature, 2015, 521(7553): 436-444.
[20]
Esteva A, Kuprel B, Novoa RA, et al. Dermatologist-level classification of skin cancer with deep neural networks (vol 542, pg 115, 2017)[J]. Nature, 2017, 546(7660): 686-686.
[21]
Bejnordi EB, Veta M, Van Diest JP, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer[J]. JAMA, 2017, 318(22): 2199-2210.
[22]
张悦，初春燕，余双，等. 人工智能应用于青光眼临床筛查及卫生效益分析[J]. 现代生物医学进展，2020，20(10): 1868-1872.
[23]
张悦，余双，马锴，等. 人工智能关于视盘区多任务深度学习模型在青光眼分类中的应用[J/CD]. 中华眼科医学杂志(电子版)，2020，10(2): 70-75.
[24]
Ko YC, Wey SY, Chen WT, et al. Deep learning assisted detection of glaucomatous optic neuropathy and potential designs for a generalizable model[J]. PLoS One, 2020, 15(5): e0233079.
[25]
Ting DSW, Pasquale LR, Peng L, et al. Artificial intelligence and deep learning in ophthalmology[J]. Br J Ophthalmol, 2019, 103(2): 167-175.
[26]
Yoo TK, Choi JY, Seo JG, et al. The possibility of the combination of OCT and fundus images for improving the diagnostic accuracy of deep learning for age-related macular degeneration: a preliminary experiment[J]. Med Biol Eng Comput, 2019, 57(3): 677-687.
[27]
Lee J, Kim YK, Park KH, et al. Diagnosing glaucoma with spectral-domain optical coherence tomography using deep learning classifier[J]. J Glaucoma, 2020, 29(4): 287-294.
[28]
Asaoka R, Murata H, Iwase A, et al. Detecting Preperimetric glaucoma with standard automated perimetry using a deep learning classifier[J]. Ophthalmology, 2016, 123(9): 1974-1980.
[29]
Muhammad H, Fuchs T, De Cuir N, et al. Hybrid deep learning on single wide-field optical coherence tomography scans accurately classifies glaucoma suspects[J]. 2017, 26(12): 1086-1094.
[30]
Ahn JM, Kim S, Ahn KS, et al. A deep learning model for the detection of both advanced and early glaucoma using fundus photography[J]. PLoS One, 2018, 13(11): e0207982.
[31]
林浩添，吴晓航. 加快基于眼科图像数据库的眼病人工智能辅助诊断平台建设[J]. 中华实验眼科杂志，2018，36(8): 577-580.
[1] 张梅芳, 谭莹, 朱巧珍, 温昕, 袁鹰, 秦越, 郭洪波, 侯伶秀, 黄文兰, 彭桂艳, 李胜利. 早孕期胎儿头臀长正中矢状切面超声图像的人工智能质控研究[J]. 中华医学超声杂志（电子版）, 2023, 20(09): 945-950.
[2] 杨菲菲, 林锡祥, 陈亦新, 王秋霜, 张丽伟, 陈煦, 张梅青, 王淑华, 何昆仑. 基于深度学习的超声心动图自动识别节段性室壁运动异常的研究[J]. 中华医学超声杂志（电子版）, 2023, 20(04): 424-429.
[3] 唐玮, 何融泉, 黄素宁. 深度学习在乳腺癌影像诊疗和预后预测中的应用[J]. 中华乳腺病杂志(电子版), 2023, 17(06): 323-328.
[4] 范帅华, 郭伟, 郭军. 基于机器学习的决策树算法在血流感染预后预测中应用现状及展望[J]. 中华实验和临床感染病杂志(电子版), 2023, 17(05): 289-293.
[5] 李晓阳, 刘柏隆, 周祥福. 大数据及人工智能对女性盆底功能障碍性疾病的诊断及风险预测[J]. 中华腔镜泌尿外科杂志(电子版), 2023, 17(06): 549-552.
[6] 胡博文, 戴英波. 泌尿外科机器人手术新趋势[J]. 中华腔镜泌尿外科杂志(电子版), 2023, 17(04): 418-421.
[7] 李腾成, 狄金明. 2023 V1版前列腺癌NCCN指南更新要点解读[J]. 中华腔镜泌尿外科杂志(电子版), 2023, 17(04): 313-318.
[8] 邢晓伟, 刘雨辰, 赵冰, 王明刚. 基于术前腹部CT的卷积神经网络对腹壁切口疝术后复发预测价值[J]. 中华疝和腹壁外科杂志(电子版), 2023, 17(06): 677-681.
[9] 邢晓伟, 刘雨辰, 王明刚. 人工智能技术在疝和腹壁外科领域的应用及展望[J]. 中华疝和腹壁外科杂志(电子版), 2023, 17(04): 390-393.
[10] 韩冰, 顾劲扬. 深度学习神经网络在肝癌诊疗中的研究及应用前景[J]. 中华肝脏外科手术学电子杂志, 2023, 12(05): 480-485.
[11] 雷漫诗, 邓锶锶, 汪昕蓉, 黄锦彬, 向青, 熊安妮, 孟占鳌. 人工智能辅助压缩感知技术在上腹部T2WI压脂序列中的应用[J]. 中华肝脏外科手术学电子杂志, 2023, 12(05): 551-556.
[12] 葛云鹏, 崔红元, 宋京海. 人工智能在原发性肝癌诊断、治疗及预后中的应用[J]. 中华肝脏外科手术学电子杂志, 2023, 12(04): 367-371.
[13] 王晓东, 汪恺, 葛昭, 丁忠祥, 徐骁. 计算机视觉技术在肝癌肝移植疗效提升中的研究进展[J]. 中华肝脏外科手术学电子杂志, 2023, 12(04): 361-366.
[14] 李京珂, 张妍春, 武佳懿, 任秀瑜. 深度学习在糖尿病视网膜病变筛查、评级及管理中的研究进展[J]. 中华眼科医学杂志(电子版), 2023, 13(04): 241-246.
[15] 胡平, 鄢腾峰, 周海柱, 祝新根. 人工智能在非增强CT图像中颅内出血早期检出和血肿分割的研究进展[J]. 中华脑血管病杂志(电子版), 2023, 17(04): 410-416.
阅读次数
全文


摘要

版权所有 中华医学会 中华医学电子音像出版社有限责任公司  京ICP备14006079号-1  网络出版服务许可证：(署)网出证(京)字第075号