A DEEP LEARNING MODEL WITH SMART PADDING FOR EARLY RETINAL DISEASE DIAGNOSIS

The project introduces an AI-powered automated diagnosis system designed to accurately and efficiently detect eye diseases from retinal images.By combining advanced deep learning methodologies with an intuitive graphical user interface, the system provides clinicians and researchers with a reliable tool for early disease detection and in-depth analysis.The application begins with robust dataset management capabilities, allowing users to upload retinal images and categorize them into specific diagnostic classes, including Diabetic Retinopathy (DR), Macular Hole (MH), Normal, and Other Diseases/Conditions (ODC).Essential preprocessing steps such as image resizing, normalization, and advanced data augmentation are incorporated to increase dataset diversity and enhance model generalization.At the core of the system are two deep learning model architectures.The first is a pre-existing Deep Neural Network (DNN) trained using the Stochastic Gradient Descent (SGD) optimizer.This model is composed of multiple convolutional layers, batch normalization, pooling, and dense layers, enabling it to effectively extract and learn complex features from retinal images.The second model is a newly proposed Convolutional Neural Network (CNN) utilizing the Adam optimizer with a specific configuration of valid padding (AVP).This architecture incorporates additional dropout layers and batch normalization to reduce overfitting and improve generalization capabilities.Both models are rigorously evaluated using key performance metrics such as accuracy, precision, recall, and F1-score.Detailed classification reports and confusion matrices offer comprehensive insights into the models' diagnostic performance.Experimental results show that while both models perform reliably, the proposed CNN with AVP outperforms the DNN in terms of diagnostic accuracy and the ability to distinguish between subtle disease features.The system's modular architecture and extensive use of data augmentation make it adaptable to various clinical datasets and scalable for real-world implementation.This research not only demonstrates the effectiveness of deep learning in the medical imaging domain but also delivers a practical, userfocused solution for automated retinal disease diagnosis contributing significantly to the development of next-generation computer-aided diagnostic systems.