• Early Detection of Diabetic Retinopathy Using EfficientNet-Based Convolutional Neural Networks

Diabetic retinopathy (DR), a diabetes-related eye disease, affects 27% of diabetic individuals globally, causing approximately 0.4 million blindness cases annually, with prevalence nearing 40% in developing regions. Early detection is challenging due to subtle symptoms and limited access to skilled ophthalmologists, particularly in underserved areas. Existing diagnostic models suffer from limited datasets, poor generalization, and integration challenges with medical systems. This project addresses these issues by developing an EfficientNet-B0-based deep learning model to detect and classify DR into five severity levels (No DR, Mild, Moderate, Severe, Proliferative) using a balanced Kaggle dataset of 10,000 retinal images. Pre-processing techniques, including Contrast Limited Adaptive Histogram Equalization (CLAHE), resizing, and data augmentation, optimized image quality, while transfer learning and attention mechanisms enhanced feature extraction and model sensitivity. The model achieved an accuracy of 94.97%, sensitivity of 94.84%, and specificity of 95.1%, outperforming traditional methods. Integrated into a Flask-based web application with Grad-CAM visualization, the model provides an interpretable, user-friendly interface for physicians, enabling scalable and cost-effective screening. This approach ensures early detection, reducing vision loss risks and improving patient outcomes in resource-limited settings. By combining advanced deep learning with practical deployment, this work advances medical imaging applications and supports public health initiatives for accessible DR screening. Future work includes expanding the dataset with diverse clinical images and leveraging Tensor Processing Units to enhance training efficiency and model robustness.