Classification of optic disc shape in glaucoma using machine learning based on quantified ocular parameters.
Classification of optic disc shape in glaucoma using machine learning based on quantified ocular parameters.
复制标题
DOI：
10.1371/journal.pone.0190012
复制
发表时间：
2017
期刊：
影响因子：
3.7
通讯作者：
Nakazawa T
中科院分区：
文献类型：
--
作者：
Omodaka K;An G;Tsuda S;Shiga Y;Takada N;Kikawa T;Takahashi H;Yokota H;Akiba M;Nakazawa T
本研究旨在开发一种基于机器学习的算法，利用眼科检查仪器获得的定量参数对开角型青光眼（OAG）患者的视盘进行客观分类。本研究纳入105例OAG患者163只眼（年龄：62.3±12.6，Humphrey field分析仪平均偏差：-8.9±7.5 dB）。3名青光眼专家将患者的眼睛分为Nicolela的4种视盘类型。随机选取114只眼作为训练数据，49只眼作为测试数据。利用训练数据对神经网络进行训练，并用测试数据对神经网络进行评估。我们利用7个患者背景特征、48个量化OCT（扫描源OCT； DRI OCT Atlantis, Topcon）值，包括视盘形貌和乳头状视网膜神经纤维层厚度（cpRNFLT）、36个激光散斑血流成像的血流参数等91种定量数据，构建机器学习分类模型。为了从91个参数中提取重要特征，使用了最小冗余、最大相关性和遗传特征选择。针对NN测试数据的验证准确率为87.8% （Cohen’s Kappa = 0.83）。神经网络的重要特征是水平盘角度、球等效、杯面积、年龄、6扇形颞上cpRNFLT、平均杯深、平均鼻缘盘比、最大杯深和上象限cpRNFLT。所提出的机器学习系统已被证明是不同类型光盘的良好标识符，具有较高的准确性。此外，这里报告的计算置信水平应该对OAG护理非常有帮助。
This study aimed to develop a machine learning-based algorithm for objective classification of the optic disc in patients with open-angle glaucoma (OAG), using quantitative parameters obtained from ophthalmic examination instruments. This study enrolled 163 eyes of 105 OAG patients (age: 62.3 ± 12.6, mean deviation of Humphrey field analyzer: -8.9 ± 7.5 dB). The eyes were classified into Nicolela’s 4 optic disc types by 3 glaucoma specialists. Randomly, 114 eyes were selected for training data and 49 for test data. A neural network (NN) was trained with the training data and evaluated with the test data. We used 91 types of quantitative data, including 7 patient background characteristics, 48 quantified OCT (swept-source OCT; DRI OCT Atlantis, Topcon) values, including optic disc topography and circumpapillary retinal nerve fiber layer thickness (cpRNFLT), and 36 blood flow parameters from laser speckle flowgraphy, to build the machine learning classification model. To extract the important features among 91 parameters, minimum redundancy maximum relevance and a genetic feature selection were used. The validated accuracy against test data for the NN was 87.8% (Cohen’s Kappa = 0.83). The important features in the NN were horizontal disc angle, spherical equivalent, cup area, age, 6-sector superotemporal cpRNFLT, average cup depth, average nasal rim disc ratio, maximum cup depth, and superior-quadrant cpRNFLT. The proposed machine learning system has proved to be good identifiers for different disc types with high accuracy. Additionally, the calculated confidence levels reported here should be very helpful for OAG care.
登录
查看更多内容
影响因子：
13.7
作者：
Leske, M. Cristina;Heijl, Anders;Yang, Zhongming
通讯作者：
Yang, Zhongming
影响因子：
3.7
作者：
Meng, Nana;Zhang, Ping;Qu, Yi
通讯作者：
Qu, Yi
影响因子：
13.7
作者：
Nicolela, MT;McCormick, TA;Chauhan, BC
通讯作者：
Chauhan, BC
影响因子：
13.7
作者：
Nicolela, MT;Drance, SM
通讯作者：
Drance, SM
影响因子：
4.4
作者：
Sawada, Yu;Hangai, Masanori;Yoshitomi, Takeshi
通讯作者：
Yoshitomi, Takeshi