Detection of glaucoma from fundus images with deep learningtechniques

Today, the rapid increase in the human population has caused an increase in the workload in the field of healthcare. Early diagnosis, diagnosis and treatment methods of diseases have become extremely important for doctors and patients. Identifying physiological changes in the human eye is a challenging task in the healthcare industry. For this reason, retinal image analysis has attracted more attention among researchers in disease screening and diagnosis in disease detection systems today. Early diagnosis of diseases that are especially difficult to detect is important for the continuity of human life. Glaucoma is an eye disease that is basically caused by the deterioration of the optic nerve in the eye. Early detection of this disease, which does not show any symptoms in terms of definitive diagnosis in its initial stage, is very difficult. In this study, a method for detecting glaucoma disease from publicly accessible retinal fundus images using a deep learning approach combined with a metaheuristic optimization algorithm is proposed. In the proposed method, feature extraction was first made from the dataset consisting of retinal fundus images with DenseNet201 and Inceptionv3 architectures, which are Convolutional Neural Network architectures. Whale Optimization Algorithm (BOA) and Particle Swarm Optimization (PSO), metaheuristic optimization algorithms, were used to eliminate irrelevant features from the dataset and reach ideal data. BOA and PSO were used in optimal feature selection according to different population sizes and iteration numbers. After the original data set and feature selection were made, Naive Bayes, Support Vector Machine, K-Nearest Neighbor, Decision Tree, Logistic Regression, and Ensemble Learning classifiers were applied to the new data sets obtained. Accuracy, sensitivity, specificity, and f1-criterion metrics were used to evaluate the performance of the classification algorithms. According to the results obtained, it was observed that promising values were obtained in the performance of the classifier from the sub-datasets determined by BOA and PSO.