Glaucoma Diagnosis: Handcrafted Features Versus Deep Learning

Abstract

Glaucoma is one of major cause of visual impairment. Glaucoma is caused by increased fluid pressure on the optic nerve leading to permanent blindness. Timely detection is crucial to prevent loss of vision. We propose, automated Glaucoma diagnosis using textural features and SVM classifier. The approach uses two textural features namely, Local Binary Pattern (LBP) and Gray Level Co-occurrence Matrix (GLCM) based features. In this study, the distortions caused in the optic nerve, optic disc along with the entire retina background are taken into account to extract the texture of the eye. The combined feature set of LBP and GLCM is then used by the Support Vector Machine (SVM) classifier to classify the images into glaucomatous and non-glaucomatous classes. The method is implemented and tested for the Medical Image Analysis Group (MIAG) Retinal Image database. Accuracy of 94% is obtained along with precision, recall, F1 score being 0.9375, 0.9473, 0.9424 respectively. The area under the ROC curve was found to be 94%. The results obtained are compared with deep learning based ResNet model. The accuracy obtained with ResNet is 92.63%, which shows that textural features with SVM classifier is an effective method for automated glaucoma detection.