Automated Multi-Class Retinal Disease Detection Using Deep Convolutional Neural Networks And Transfer Learning

Abstract

Early detection of ocular diseases is essential for preventing progressive visual impairment and reducing the burden of avoidable blindness. However, conventional retinal screening depends heavily on trained ophthalmologists and manual interpretation of retinal images, which can limit the scalability and accessibility of screening services. This study presents an automated deep-learning framework for the multi-class classification of retinal fundus images into four categories: cataract, diabetic retinopathy, glaucoma, and normal. The proposed model employs transfer learning using a pre-trained EfficientNetB3 convolutional neural network as the feature-extraction backbone. To improve disease-specific feature learning while limiting overfitting, the deeper layers of the backbone were fine-tuned and combined with global average pooling, batch normalization, dropout regularization, and fully connected classification layers. The retinal image dataset was screened for corrupted files, divided into training, validation, and test partitions, and subjected to training-specific augmentation using horizontal flipping, rotation, zooming, and contrast variation. Model performance was evaluated using accuracy, precision, recall, and F1-score. The proposed classifier achieved a test accuracy of 99.58%, demonstrating strong classification performance across the four retinal categories. The trained model was further integrated into a full-stack web application consisting of a React-based frontend and Flask-based backend. A pre-inference image-validation mechanism was incorporated to identify unsuitable, non-retinal, or severely degraded image inputs before classification. The developed framework demonstrates the potential of transfer-learning-based deep neural networks for scalable and accessible preliminary retinal disease screening. Nevertheless, independent external and prospective clinical validation is required before the system can be considered suitable for real-world clinical use.