Abstract:
Detection of retinal diseases early is crucial to preventing permanent vision loss. However, traditional diagnostic methods often depend on specialists manually interpreting results, which can be both time-consuming and inconsistent. This research introduces a deep learning framework designed to automate the classification of early-stage retinal diseases using color fundus images. By utilizing convolutional neural networks (CNNs), the study examines transfer learning models that were trained on datasets from the University of Gondar Referral Hospital Eye Clinic. It also incorporates preprocessing techniques like resizing, cropping, and augmentation to enhance clinically relevant features while minimizing noise. Through extensive experimentation, the model's performance was evaluated across major retinal conditions, including diabetic retinopathy, glaucoma, and age-related macular degeneration. The quantitative analysis, which employed metrics such as accuracy, F1-score, and area precision, revealed that the deep neural network achieved impressive diagnostic performance, showing remarkable resilience against inter-patient variability. The results indicate that the deep learning model DensNet201 can be a powerful tool for early retinal screening, boasting an overall accuracy of 92.78%. Additionally, it showed high precision, recall, and F1-scores across all categories, especially for normal and glaucoma cases. The confusion matrix further supported these findings, indicating that misclassifications were minimal and mostly occurred between diabetic retinopathy and other classes with similar visual characteristics. Future research will aim to enhance interpretability and facilitate real-time integration into portable screening devices.
Keywords: Deep Learning, Convolutional Neural Networks (CNN), Retinal Disease, Fundus Images, Medical Image Classification, Diabetic Retinopathy, Glaucoma, Age-related Macular Degeneration, Early Diagnosis, Transfer Learning, Automated Screening, Ophthalmology.
