Glaucoma Detection from Retinal Images using Generic Features

Abstract

Glaucoma is a silent disease that, left untreated, causes severe visual impairments which can progress to irreversible blindness. Fortunately, early detection and proper treatment can control the development of glaucoma and in turn limit further progression of associated visual impairments. However, periodic manual diagnosis of glaucoma necessary for its early diagnosis would require abundancy of experts, besides being invasive, expensive, and time consuming. Computer aided diagnosis (CAD) can thus serve as a game changer in the early detection of glaucoma by bringing clinician to the level of an expert. Moreover, CAD has the advantages of being non-invasive, simple, and cost effective. In this work, an automated generic glaucoma detection algorithm is presented in which statistical and textural features are computed from the optic nerve head (ONH) region within retinal images. Several analyses are performed to compare glaucoma classification performance considering different contrast enhancement techniques (histogram equalization - contrast limited adaptive histogram equalization) and color models (RGB - HSV - CIELAB). Feature selection is then used to find the best set of features for each of the different experiments. Best performance was achieved when textural features were computed from the histogram equalized CIELAB channels, resulting in an accuracy of 92.5%, sensitivity of 95.0%, and specificity of 90.0% considering a public dataset consisting of 40 glaucomatous and 40 healthy retinal images.