Early Glaucoma Detection Using Deep Neural Networks and Adaptive Neuro Fuzzy Logic

Abstract

Glaucoma is the second most common cause of vision loss around the world, after cataracts. However, unlike cataracts, glaucoma usually can't be fixed. The number of people with glaucoma is expected to rise from 64 million in 2015 to 76 million in 2020 and 111 million in 2040. Finding early signs of glaucoma is still hard for eye doctors in developing countries. In real life, there is always a gap between early diagnosis of a disease and a patient's medical history. This makes it hard to classify people correctly. Most of the time, there are four main problems with deep learning algorithms for detecting glaucoma. These are a limited dataset, significant coefficients of the statistical data, measuring intraocular pressure, and CDR. In recent studies, researchers have tried to improve the accuracy of classification by using different combinations of features, statistical data, or changes to deep learning algorithms. Using these optimal feature selection algorithms raises the cost of computation. The goal of this study was to find glaucoma early and improve the accuracy of classification. This was done with the help of Fundus Imaging and a three-headed fusion of Convolutional Autoencoding Classifier Framework, Adaptive Neuro Fuzzy logic, and statistical analysis using wavelet transformation. These techniques help ophthalmologists diagnose and watch for eye diseases by giving them clear and accurate information about the structure of the optic nerve head. The proposed model network is optimized so that the image reconstruction error, the classification error based on a task learning procedure, and the fuzzy logic glaucoma predictor based on the statistical data model are all as small as possible. This article also gives a review and analysis of the work that came before. Its accuracy is shown by getting an accuracy of 95.89% on the Kin Hospital early glaucoma dataset. This shows that the architecture is new and accurate. Based on the proposed clinical inputs, the ANFIS results come out to be 95.4%.