<![CDATA[Visual impairment from diabetic retinopathy, glaucoma, and cataracts remains a critical global health issue, emphasizing the need for early and accurate diagnosis to prevent permanent vision loss. This research presents an automated detection system utilizing ResNet-50, a deep learning architecture, to classify fundus images into multiple retinal disease categories. Unlike conventional convolutional neural networks used in prior studies, this approach leverages ResNet-50's residual learning mechanism to better identify complex retinal patterns. The study employed 4,184 fundus photographs from Kaggle, divided into four classes: cataract, diabetic retinopathy, glaucoma, and normal. Images were preprocessed through resizing to 224×224 pixels, normalized with ImageNet parameters, and augmented using random rotation and flipping techniques to enhance model generalization. Dataset splitting followed stratified sampling with an 80-20 train-test ratio, maintaining balanced class representation. Model training spanned 20 epochs using the Adam optimizer across three learning rates: 0.1, 0.01, and 0.001. The 0.001 learning rate produced optimal results with 90.35% accuracy, 90.28% precision, 90.18% recall, and 90.21% F1-score. The confusion matrix indicated strong performance in detecting diabetic retinopathy (219 correct predictions) and normal cases (189 correct predictions), though minor misclassifications occurred between glaucoma and normal categories. These findings validate ResNet-50's residual architecture as an effective tool for extracting discriminative retinal features, offering a computationally efficient solution for automated eye disease screening. Future work should incorporate explainability methods like Grad-CAM to enhance clinical interpretability and build trust among healthcare professionals in AI-assisted diagnostic systems.]]>
Copyrights © 2026
