Diabetic retinopathy (DR) is one of the leading causes of preventable blindness worldwide, highlighting the critical need for accurate and efficient automated screening systems. Most existing deep learning-based DR grading approaches rely on supervised classification of single-timepoint fundus images and do not explicitly account for disease progression or the ordinal nature of DR severity levels. In this thesis, a novel Progression-Aware Contrastive Learning framework is proposed to improve representation learning for diabetic retinopathy grading. The proposed framework employs an anchor-based contrastive learning strategy that constructs progression-aware image pairs using severity labels and ordinal distance information. An ordinal-distance-weighted contrastive loss is introduced to implicitly capture clinically meaningful progression patterns without requiring explicit longitudinal data. This design enables the learning of more discriminative and severity-consistent feature representations while maintaining computational efficiency suitable for large-scale screening applications. Experimental evaluations on benchmark diabetic retinopathy datasets demonstrate that the proposed framework achieves an improvement of approximately 2–3% in classification accuracy and 0.03–0.05 in quadratic weighted kappa (QWK) compared to representative efficient baseline methods. Qualitative analyses further indicate that proposed framework substantially reduces severe misclassifications between distant DR grades, enhancing the clinical reliability of automated grading. Overall, the results confirm that progression-aware contrastive learning offers an effective and practical solution for improving diabetic retinopathy screening systems.
