Explainable Deep Learning Models for Clinical Decision Support in Diabetic Retinopathy Detection

This study explores an integrated approach to diabetic retinopathy (DR) detection by combining advanced image preprocessing techniques with transfer learning. Preprocessing methods such as vignetting reduction, denoising, adaptive contrast equalization, and color normalization are employed to enhance the quality and consistency of retinal images. Pre-trained deep learning models, including AlexNet, GoogLeNet, ResNet, and VGGNet, are leveraged for feature extraction and classification. The methodology addresses critical challenges such as feature mismatches and information loss by fine-tuning extracted features to improve model adaptability and diagnostic accuracy. Performance evaluation focuses on metrics like confusion matrices, computational efficiency, and resource requirements, ensuring a robust and clinically applicable framework. This approach demonstrates the potential of combining preprocessing with transfer learning to advance automated DR screening, reduce diagnostic disparities, and improve ophthalmic care outcomes.