Abstract:
Diabetic retinopathy (DR) is a leading cause of preventable blindness worldwide. Automated
DR grading is inherently challenging due to the ordinal nature of disease stages, where
misclassifications across distant stages (e.g., No DR vs. Severe DR) are more critical than
between adjacent ones. Existing contrastive and classification-based methods often fail to model
these ordinal relationships between severity stages, leading to clinically inconsistent grading
outcomes. To address this, we propose an Order-aware Contrastive Loss (OCL) that incorporates
inter-class distance weighting to enforce ordinal consistency during representation learning.
Furthermore, we introduce a multi-branch architecture built upon a DenseNet-121 backbone,
which integrates OCL with two complementary objectives in a unified multi-head framework. The
entire network is jointly optimized to enhance both predictive accuracy and ordinal alignment. To
the best of our knowledge, this is the first contrastive learning framework that explicitly integrates
ordinal distance weighting for diabetic retinopathy grading. The proposed approach is evaluated
on two benchmark datasets: EyePACS, a large-scale dataset containing 35,126 training and 42,670
test images across five stages, and APTOS2019, which includes 3,662 images. On EyePACS,
our method achieves 86.93% accuracy, 0.8644 quadratic weighted kappa (QWK), and 0.8694 F1-
score, surpassing the previous state-of-the-art (84.55% accuracy, 0.8631 QWK, 0.8440 F1-score).
Similar improvements are observed on APTOS2019. These results highlight the effectiveness
of incorporating order-aware representation learning with multi-objective optimization on a
strong DenseNet based architecture, setting a new benchmark for ordinal DR classification.
