Complex Texture Features for Glaucomatous Image classification System using Fundus Images
Abstract
In this paper, an efficient approach for glaucomatous image classification system using fundus images is proposed. The main aim of this study is to detect glaucoma accurately in order to reduce the visual loss and impairment. The proposed system uses two important texture features; Gray Level Co-occurrence Matrix (GLCM) and Local Binary Pattern (LBP) in an efficient manner. These texture features are extracted not only from the fundus image but also the optical density image obtained from the fundus image. Before extracting features, region of interest is obtained from the Green channel of the fundus image as it has high contrast than other two colour components. Support Vector Machine (SVM) classifier is used for the classification of fundus image into normal or abnormal based on the extracted features. Results show that the proposed system provides promising results with 100% sensitivity and 99% specificity.
Keywords
Download Options
Introduction
Blindness is the lack of visual perception due to neurological or physiological factors. The major causes of blindness are glaucoma, cataract, age-related macular degeneration, and corneal opacity. Among them, glaucoma is one of the irreversible blindness. An extensive literature survey has been done and some of them are outlined here. An approach to detect glaucoma using Cup to Disc Ratio (CDR) and ISNT rule is discussed in [1]. At first, Region Of Interest (ROI) is extracted and CDR is measured by segmenting optic disc and cup region. The blood vessels in the optic disc area are tracked using hessian based vessel enhancement technique to compute ISNT ratio. Then, SVM classifier is used for classification.
Deep convolutional neural network based glaucoma detection is discussed in [2]. It consists of six layers; four convolutional layers and two fully connected layers. Overlapping-pooling layers and response-normalization layers are adopted to reduce the overfitting problem. A review of various automated techniques for glaucoma diagnosis is discussed in [3] including active contour model, super pixel clustering, vessel bend, simple linear iterative clustering, and pallor information.
Four different features such as CDR, horizontal to vertical CDR, cup to disc area ratio and rim to disc area ratio are used for glaucoma diagnosis in [4]. Optic disc segmentation is done with geodesic active contour model and cup segmentation is based on pallor appearance in the optic disc region. Finally, naïve Bayes, K- nearest neighbour and SVM classifiers are used for classification. Image based features along with segmentation based features are used for glaucoma diagnosis in [5]. Illumination correction is done before extracting features. Optic disc region is detected using Hough transform and template matching is used for ROI extraction. Finally, SVM classifier is adopted for classification.
Conclusion
An efficient fundus image classification system is presented in this paper. The proposed approach utilizes GLCM and LBP features extracted from the original image as well as from their optical density images. SVM classifier is designed for the classification of given fundus image into normal or glaucomatous image. The performance of GLCM, LBP and hybrid features are analyzed using the same set of training and testing fundus images. It is observed that the proposed system provides 99% classification accuracy while using the hybrid features. Also, it is observed that 100% sensitivity and 99% specificity is obtained by the proposed system.