Transfer Learning and Advanced Image Analysis for Diabetic Retinopathy Staging

Diabetic retinopathy (DR) is a critical ophthalmological complication that can lead to vision loss if not diagnosed early. This study explores the automated detection and the classification of DR severity by using deep learning techniques by investigating five pre-trained convolutional neural networks—AlexNet, DenseNet121, ResNet50, InceptionV3, and EfficientNetB0 which were trained on the ImageNet and then evaluated on the gaussian-filtered retinal images across five different stages. The models were assessed using k-fold cross-validation and comprehensive evaluation metrics, including accuracy, Euclidean distance, and cosine similarity. Furthermore, explainable AI tools like LIME (Local Interpretable Model-Agnostic Explanations) were deployed in order to increase the model’s transparency by visually highlighting the retinal regions that are most responsible for model decisions, thereby supporting clinical interpretability as traditional deep learning models are generally black-box models which offers very less transparency. DenseNet121 and InceptionV3 display the highest validation accuracy of about 74%, with a strong generalization performance and consistent focus on the clinically relevant regions in retinal images. ResNet50 and EfficientNetB0 showed a lower validation performance of approximately 56.6 and 59.6% respectively, indicating potential overfitting and reduced generalization. These results underscore the effectiveness of transfer learning for analyzing medical images and highlight deep learning’s potential role in assisting early DR detection and stage classification.