A CNN-BASED FRAMEWORK FOR EFFICIENT DETECTION OF EYE DISEASE IN FUNDUS IMAGES

Abstract

Early diagnosis and treatment of eye diseases (ED) are essential for improving patient outcomes and preventing permanent vision loss. Retinal fundus image screening is a widely used technique for the preliminary assessment of ocular conditions; however, manual interpretation of these images is time-consuming and requires expert knowledge. Recent advances in deep learning (DL) have shown significant potential in automating medical image analysis, particularly for retinal disease detection. Despite achieving strong performance, further improvements can be made by integrating effective preprocessing techniques and optimized model architectures.

This study proposes an automated framework for early-stage detection and classification of eye diseases using retinal fundus images and a custom-designed convolutional neural network (CNN). The methodology consists of several key stages. First, fundus images are preprocessed using multiple image enhancement techniques, including resizing, green channel extraction to emphasize retinal structures, min-max normalization to standardize pixel intensity, and data augmentation to increase dataset diversity and improve model generalization. The processed images are then labeled into four classes—Normal, Cataract, Glaucoma, and Diabetic Retinopathy—and split into training and testing datasets. A custom CNN architecture is developed, comprising multiple convolutional layers for feature extraction, max-pooling layers for dimensionality reduction, a flatten layer for feature transformation, and fully connected dense layers with dropout regularization to prevent overfitting. The final output layer uses a SoftMax activation function to perform multi-class classification. The model is implemented using Python on the Google Collab platform with standard deep learning and computer vision libraries. The proposed model is evaluated on a publicly available Kaggle dataset containing 4,217 retinal images, including 1,074 normal, 1,038 cataract, 1,007 glaucoma, and 1,098 diabetic retinopathy cases. Experimental results demonstrate strong performance, achieving 96.8% training accuracy and 94.2% testing accuracy, with a sensitivity of 93.5%, specificity of 95.1%, and a low loss of 0.18. Class-wise evaluation further shows high precision and recall across all categories, with values exceeding 90%, indicating reliable and consistent classification performance.

Overall, the proposed framework effectively automates the detection of retinal diseases and reduces reliance on manual diagnosis. It offers a promising tool for assisting ophthalmologists in early detection and clinical decision-making