KMoCoL: k-Positive Momentum Contrastive Learning for Imbalanced Diabetic Retinopathy Grading

Diabetic retinopathy (DR), a leading cause of preventable visual impairment, necessitates early screening for effective intervention. Deep learning-based DR grading approaches, often formulated as image classification tasks, have shown promise using end-to-end supervised models. However, clinical datasets inherently exhibit class imbalance, which often leads to end-to-end models favoring the majority class and hindering the accurate classification of minority classes. To address the class imbalance challenge in DR grading, this article proposes a two-stage k-positive momentum contrastive learning (KMoCoL) framework. KMoCoL leverages contrastive learning to generate balanced and informative features, while incorporating label information for enhanced semantic discrimination. The proposed approach effectively mitigates class imbalance and improves the overall accuracy of DR grading, achieving state-of-the-art performance on two benchmark DR grading datasets. Furthermore, analysis of the experimental results reveals that KMoCoL also contributes to reducing the effects of interclass correlation.