An Automatic Screening Method For Detecting Glaucoma Using Multi-class Support Vector Machine P. Vejjanugraha 1 , W. Kongprawechnon 1 , T. Kondo 1 , K. Tungpimolrut 2 , H. Kaneko 3 and S. Sintuwong 4 1 School of Information, Computer, and Communication Technology, Sirindhorn International Institute of Technology, Thammasat University, Thailand. 2 National Electronics and Computer Technology Center, Thailand. 3 Tokyo Institute of Technology, Japan. 4 Mettapracharak Eye Center, Mettapracharak Hospital, Thailand. (E-mail: pikulvej@gmail.com, waree@siit.tu.ac.th, and tkondo@siit.tu.ac.th) Abstract—This paper introduces an automatic screening method for detecting the development of glaucoma using multi- class support vector machine (Multi-class SVM). Due to the high rate of blindness worldwide from glaucoma, the early detection is a must for early treatment and preserving the eye sight. Only the diagnoses of the optic cup and the optic disc from the ophthalmologists or technicians are subjective. The multi-class SVM is introduced to overcome this problem by classifying it into 3 classes; normal case, glaucoma suspect case, and glaucoma case. This research proposes 2 types of multi-class SVM; One-vs- the-rest SVM and SVM with 2-stage of decision tree. The linear and polynomial kernel functions are selected to generate the classifier. The evaluation is done by 10-fold cross validation technique which does not only focus on the high accuracy rate, but also considers the number of false negative rate. The final results show that SVM with 2-stage decision tree with polynomial kernel function provides the best performance to classify each stage of glaucoma. Keywords—Glaucoma, Cup-to-disc ratio, image segmentation, Feature extraction, and Multi-class support vector machine. I. INTRODUCTION Glaucoma is recognized chronic degenerative health problem worldwide, estimated to affect over 60 million people of all ages [1]. It is asymptomatic in the early stage. The patients stay unnoticed until the vision field is lost gradually. Glaucoma can be computerized based on the peripapillary chorioretinal atrophy (PPA), Retinal nerve fiber layer defect (NFLD), vision field test, and the optic cup-to-disc ratio (CDR) [2-3]. When glaucoma damages the optic nerve or the nerve fiber layer around the optic nerve head (ONH), it cannot be cured. Therefore, the early detection could bring the early treatment to prevent the vision loss. The primary treatment is to reduce the intraocular pressure (IOP) inside the eyes by drops. The other clinical factors are ages (over 40), glaucoma in the family, diabetes mellitus [4]. Nowadays, to analyze and diagnose glaucoma, the followings are generally considered; the clinical data, optical coherence tomography (OCT) scan, and the vision field examination. However, in some rural area, the lack of ophthalmologists and effective equipment is appeared. The analyzing result of inexperienced trainee might cause some mistakes and that probably make some damages or the loss of vision to the patients. To overcome the subjective results, the learning machine is introduced to classify the data basing on the training set in the data base. The feature extraction and feature selection are applied to the input fundus image. The followings are the example of the extracted features; the size optic cup and optic disc in vertical and horizontal, the area of the optic disc and optic cup, pixel intensity values, fast Fourier (FFT) coefficients, B-spline coefficients [5-6]. The classification process consists of many kinds of algorithm and theories. Mostly it is divided into 2 types; supervised and unsupervised. In this research, the supervised learning is considered. The new data will be classified based on the training set which is already labeled to be likely or unlikely. It is widely used in the real world application such as target marketing, banking analysis, agriculture, remote sensing, and medical diagnosis. Artificial neural network (ANN) is the example of traditional supervised classifier that mostly used to classify the data in binary. Moreover, this research introduces support vector machine (SVM) which is a supervised classifier that powerful to the high dimensional data. According to Ref. [7], SVM is applied to an screening system for the age-related macular degeneration (AMD). The training set consists of 4 different data sets of AMD; normal, drusen, micro aneurysms, and circinate. The features are extracted from EMD and DWT techniques. It is shown that the accuracy of EMD-SVM is higher than DWT-SVM. They conclude that features extracted from EMD techniques are more adaptive than the DWT technique. Therefore, feature selection process is very important. When the feature widely separates to the each other, the margin between classes will be maximized and provide well results. Referring to Ref. [8], when the data is imbalanced, SVM does not provide good performance, because of the weakness of soft-margins and imbalanced support vector ratio. This problem can be solved by sampling method, synthetic minority oversampling technique (SMOTE). In Ref. [9], it explains the idea to synthetic the data by an adaptive over-sampling technique based on data density (ASMOBD). The idea of ASMOBED is to synthetic the new data in the majority area and in the overlapping area in order to avoid the overfitting problem. For the outlier, it is considered as a noise and no synthetic data around that area. Moreover, SVM can be used to classify the multi-class classification. Ref. [10] is analyzed by using 2-stage SVM.