| Home  | About ScienceAsia  | Publication charge  | Advertise with us  | Subscription for printed version  | Contact us  
Editorial Board
Journal Policy
Instructions for Authors
Online submission
Author Login
Reviewer Login
Volume 43 Number 4
Volume 43 Number 3
Volume 43 Number 2
Volume 43 Number 1
Volume 43S Number 1
Volume 42 Number 6
Earlier issues
Volume 43 Number 3 Volume 43 Number 4

previous article next article 1

Research articles

ScienceAsia 43(2017): 229-239 |doi: 10.2306/scienceasia1513-1874.2017.43.229


An automatic screening method for primary open-angle glaucoma assessment using binary and multi-class support vector machines


Pikul Vejjanugrahaa,b,*, Waree Kongprawechnona, Toshiaki Kondoa, Kanokvate Tungpimolrutc, Kazunori Kotanib

 
ABSTRACT:     Glaucoma is a chronic progressive eye condition leading to permanent visual loss. An automatic screening system is necessary to detect primary open-angle glaucoma because it is an insidious disease appearing without symptoms or early warning signs. This work introduces an automatic screening technique to diagnose glaucoma using a support vector machine (SVM). Two case studies are investigated: binary-stage and multi-stage classification of glaucoma. First, there is a comparison of the performance of the hard threshold-based approach to the supervised learning approach using an SVM. Image segmentation techniques are performed to detect important features: the actual sizes of the optic cup and optic disc in vertical and horizontal directions. SVMs with a linear kernel function are used to generate the classifier model, and the results show that using threshold-based classification is inadequate to screen glaucoma. In a second case study, an SVM is applied to develop the classification algorithm focused more on the detection of the glaucoma suspect stage, which is an intermediate stage between the healthy and glaucoma stages. A polynomial kernel function is used to implement the classification model. The unbalanced decision tree (UDT) and one-versus-the-rest (OVR) techniques are combined in the models in order to overcome the limitations of an SVM. Finally, the combination of an SVM with both UDT and OVR techniques yields a reliable result with respect to belonging classes at 99.4%.

Download PDF

36 Download 57 View


a School of Information, Communication, and Computer Technologies, Sirindhorn International Institute of Technology, Thammasat University, Khlong Luang, Pathum Thani 12120 Thailand
b School of Information Science, Japan Advanced Institute of Science and Technology, Nomi, Ishikawa 923-1211 Japan
c National Electronics and Computer Technology Centre, National Science and Technology Development Agency, Khlong Luang, Pathum Thani 12120 Thailand

* Corresponding author, E-mail: pikulvej@jaist.ac.jp

Received 9 Mar 2017, Accepted 24 Sep 2017