Enhanced Diabetic Retinopathy Detection: An Explainable Semi-Supervised Approach Using Contrastive Learning

Diabetic retinopathy (DR) is a leading cause of blindness and represents a critical challenge to global vision health. Early detection is essential to preventing irreversible eye damage. Automated medical image analysis plays a pivotal role in enabling timely diagnosis. However, the development of robust diagnostic models is challenged by the scarcity of labeled data and the prevalence of imbalanced and unlabeled datasets. Semi-supervised learning offers a potential solution by leveraging unlabeled data to enhance model performance. However, it is often limited by challenges such as unreliable pseudo-labeling, the exclusion of low-confidence data, and biases introduced by imbalanced datasets. To address these limitations, we propose a novel semi-supervised learning framework for DR detection that combines similarity and contrastive learning. Our approach utilizes class prototypes and an ensemble of classifiers to generate reliable pseudo-labels for unlabeled data. Unlike traditional methods that discard unreliable samples, our framework integrates them into the training process using contrastive learning. This allows us to extract valuable features and improve overall performance. Furthermore, we enhance the model's transparency and interpretability by incorporating the explainable AI technique GradCAM, which provides insights into the model's predictions for specific images. We evaluated the proposed method on the publicly available Kaggle DR dataset for diabetic retinopathy classification. Experimental results demonstrate that our approach achieves improved performance compared to existing semi-supervised learning methods. It also effectively leverages unreliable samples, highlighting its potential to advance DR diagnosis.