An Automated Deep Learning Approach to Diagnose Glaucoma using Retinal Fundus Images

Abstract

One of the most dangerous diseases that affect the vision of the human eye is glaucoma. It is a retinal illness that damages the ONH (Optic Nerve Head) and results in everlasting blindness if it is diagnosed at an advanced stage. Permanent blindness can only be avoided through early-stage detection and treatment of glaucoma. This paper presents a model based on the architectures of convolutional neural networks that can detect glaucoma at an early stage using the fundus images. Three datasets, G1020, RIM-ONE and REFUGE, are used in the proposed model. The images are enhanced through the CLAHE (Contrast Limited Adaptive Histogram Equalization) approach and a median filter is applied to reduce the noise during preprocessing step. The variety of images is created to efficiently train the proposed model by using the data augmentation approach. Three pre-trained CNN (Convolutional Neural Network) architectures such as VGGl9, ResNet50 and EfficientNetB7 are used for the classification, and their performance is compared through different performance measures. The best classification results are achieved through the EfficieientNetB7 architecture with accuracy of 99.2%, sensitivity of 98% and specificity of 97% on G1020 dataset. The robust results obtained by the proposed method have smoothened the way towards the fast and reliable glaucoma detection system It will help the ophthalmologists and the clinicians to screen the masses and make reliable decisions about the glaucoma diagnosis.