Low-Cost Deep Learning-Based System for Early Diagnosis of Diabetic Retinopathy Using Retinal Fundus Images

Diabetic Retinopathy (DR) is a vision-threatening complication of diabetes that can lead to irreversible blindness if not detected early. However, manual screening of retinal fundus images by ophthalmologists is both timeconsuming and resource-intensive. The proposed system introduces a low-cost, deep learning-based approach for the early diagnosis and classification of DR severity levels using retinal fundus images. The system utilizes a publicly available Kaggle dataset, consisting of retinal images labeled across five severity stages (0–4), ranging from No DR to Proliferative DR. Data preprocessing involved linking images with their corresponding labels using a train.csv file and addressing class imbalance through computed class weights. A transfer learning technique based on the ResNet50 architecture pre-trained on ImageNet was applied and fine-tuned for multi-class classification. The model was trained using TensorFlow and Keras, with extensive image augmentation to enhance generalization and prevent overfitting. The trained model achieved promising accuracy and was saved for reuse. A prediction module was also developed to classify new retinal images into diagnostic categories such as “No DR,” “Moderate,” or “Severe.” The project demonstrates that a resource-efficient AI solution can support ophthalmologists in DR screening on standard computing hardware, eliminating the need for costly infrastructure and enabling scalable deployment in lowresource clinical settings