Development of classification system for eye diseases with fundus images using convolutional neural network Embedded in web and mobile applications

Abstract

This study assesses how different Convolutional Neural Network (CNN) models perform in classifying eye (ocular) disorders, focusing on glaucoma, cataracts, diabetic retinopathy, and normal cases. The models tested include ResNet-50, InceptionV3, and MobileNetV3, with particular emphasis on the lightweight and efficient MobileNetV3 due to its suitability for mobile and embedded platforms. The MobileNetV3 model achieved a training accuracy of 96% and a training loss of 19%, with a validation accuracy of 95% and a validation loss of 19% after 20 epochs. The cataract class's precision, recall, and F1-score are 97%, 100%, and 98%, respectively. For diabetic retinopathy, these metrics were 95%, 95%, and 95%; for glaucoma, they were 100%, 97%, and 98%. The InceptionV3 model for three-class classification attained a training accuracy of 98% and a training loss of 20%, with a validation accuracy of 90% and a validation loss of 35%. Although lower than the three-class model, the four-class version of InceptionV3 similarly demonstrated strong performance metrics. ResNet-50 also presented high accuracy and robust performance metrics across different classes. However, the MobileNetV3's lightweight architecture, depth-wise separable convolutions and squeeze-and-excitation modules made it suitable and deployed on web and mobile applications.