Researches Of Automatic Diagnosis Method For Keratoconus Based On Machine Learning

Google AlphaGo beat South Korean professional Go player Lee Sedol in 2016,which was a major milestone in machine learning research.From personalized recommendation songs in the music website to predictions of unknown events like weather forecast and earthquake detection,the application of machine learning in our daily life is everywhere.Machine learning has brought a big innovation to the traditional industry,and the medical field is no exception.Early application of machine learning in medical diagnosis is tumor diagnosis and brain image analysis and so on.In November 2016,Google developed an algorithm to screen early diabetic retinopathy using deep learning with far better performance than the top experts.Keratoconus is a progressive corneal disease mostly occurs in adolescence.Early patients can still correct their vision by wearing RGP,although there is a severe decreasing of vision because of rapid growth of high myopia and irregular astigmatism.In the late,rupture of the Descemet membrane makes aqueous humor come into cornea,leading to scar with permanent damage of vision and could not be corrected except corneal transplantation.The disease will be controlled and free from corneal transplantation if early keratoconus is detected.In addition,keratoconus should be excluded before a Lasik surgery to treat myopia.However,the diagnosis of early keratoconus is still difficult for doctors so that a good method for detection of keratoconus and subclinical keratoconus is urgent.In this paper,machine learning is used to diagnose keratoconus and subclinical keratoconus.Firstly,a method for collecting sample feature data and label systematically is proposed.Two-group and three-group classification experiments are developed using data from Oculyzer.The detection of keratoconus and subclinical keratoconus show high sensitivity and the sensitivity of various features for keratoconus and subclinical keratoconus detection is verified.Finally,two applications of the diagnosis method of keratoconus are given.The main contributions of this paper are in the following four aspects:(1)A method for collecting sample data systematically is presented,which includes sample feature data characterizing disease characteristics and the diagnosis of the doctor as labels.The sample labels are obtained from the electronic diagnostic system of doctors and the sample feature data was obtained from the anterior segment analyzer Oculyzer.Zernike aberration data is automatically extracted from the picture by matching with the number template.(2)The paper proposed a method of dividing the sample into normal and keratoconus with sensitivity of 99.20%and specificity of 98.26%,using SVM as model,sensitivity and specificity to measure model performance,ten-fold cross validation to test model stability.Linear kernel function and RBF kernel function are used to study the sensitivities of different features for keratoconus detection.In general,RBF kernel function shows better sensitivity than linear kernel function.The sensitivity of the features to the detection of keratoconus was all over 80%,indicating that all these features contribute to the detection of keratoconus.Corneal elevation,corneal thickness,Zernike aberration and the combination of them show the most excellent performance among the experiments.(3)A method was developed to divide the samples into three groups,normal eyes,subclinical keratoconus and keratoconus,to detect subclinical keratoconus and keratoconus.In the basic model,three two-group SVM classifiers of One-vs-All are trained to learn the characteristics in which each class is different from other two classes,and the final three-group result is determined by a new learner which integrates the results of three two-group classifier.In addition,the improved model adjusts the sample weight of the next classifier based on the error-diagnosed samples of the previous classifier,using Adaboost to enhance the two-group classifiers,and ultimately be able to detect subclinical keratoconus and keratoconus with good sensitivity.Zernike aberration and corneal elevation performance higher sensitivity in the detection of subclinical keratoconus and combination of the two features lead to further improvement.(4)The paper gives two application prospects of keratoconus diagnosis.The first is the personalized disease management system to help patients know their own condition of the illness and choose appropriate treatment.The second is the wide range of keratoconus screening in teenagers,which is not only the detection of subclinical keratoconus for adolescents,but a feedback of the keratoconus diagnosis method.