Study of feature selection for angle-closure glaucoma detection

• Main Author: Gercke, Norman.
• Other Authors: Lin Weisi
• Format: Final Year Project
• Language: English
• Published: 2013
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Pattern recognition
• Online Access: http://hdl.handle.net/10356/55089
• Institution: Nanyang Technological University

Glaucoma is an eye disease where a loss of vision occurs as a result of progressive optic nerve damage caused by high intraocular pressure within the eye. Performing feature selection on features obtained by Anterior Segment Optical Coherence Tomography (AS-OCT) images may aid in clinical diagnosis of angle-closure glaucoma. Existing feature selection algorithms such as Minimum Redundancy Maximum Relevance (MRMR) and Laplacian Score offer promising feature selection capabilities. In this project, these algorithms were used together with the AdaBoost machine learning classifier to train on a dataset provided by the National University Hospital consisting of 84 features and 156 samples split into 5 classes of anterior segment mechanisms (Iris Roll, Lens, Pupil Block, Plateau Iris and No Mechanism). The AdaBoost-MRMR MIQ method was able to produce an accuracy of 84.39% using a small set of 10 features, while the AdaBoost-Laplacian KNN Heat Kernel method was able to produce a higher accuracy of 86.66%, albeit at a higher feature set of 40 features. Similar features found in the Laplacian KNN Heat Kernel feature set were found to contribute to this accuracy despite being correlated. Low sensitivities for the iris roll class and no mechanism class in both results were suggested to be due to low sample sizes for training, and in the case of the iris roll class, due to a possible combination with other mechanisms. A larger sample size and a consideration for two or more mechanisms in a single sample could yield improved results in future works.