Diffusion-augmented contrastive learning for diabetic retinopathy severity classification

Automated diabetic retinopathy (DR) severity classification faces persistent challenges from limited annotated datasets, severe class imbalance across the five-level International Clinical Diabetic Retinopathy (ICDR) scale, and domain shift between imaging environments. This paper presents an integrated deep learning framework that extends our earlier Fused Ensemble Convolutional Inception Network (93.2% accuracy) through three complementary innovations. First, a class-conditional latent denoising diffusion probabilistic model (DDPM) generates anatomically realistic synthetic fundus images for underrepresented DR severity levels, addressing data scarcity and class imbalance simultaneously. Second, a SimCLR contrastive self-supervised pre-training protocol, applied to the combined real-and-synthetic corpus, learns discriminative retinal representations prior to supervised fine-tuning. Third, a five-model ensemble (EfficientNet-B3, ResNet-50, VGG16, VGG19, and DenseNet-121) fused via a learned Pattern-Net achieves improved generalization through architectural diversity. Evaluated on the APTOS 2019 dataset under 5-fold stratified cross-validation, the integrated framework achieves 95.1% accuracy, 0.928 quadratic weighted kappa (QWK), 0.91 F1-score, and 0.97 AUC. Class-wise improvements over the EfficientNet-B3 baseline reach +23.8 percentage points for Severe DR (61.8 → 85.6%) and +17.7 for Mild DR (71.2 → 88.9%). Independent zero-shot validation on EyePACS yields 78.4% accuracy, 0.809 QWK, 0.738 macro F1, and 0.887 macro AUC, quantifying the expected performance reduction under domain shift. Fold-level variability and paired out-of-fold statistical analyses are additionally reported, and Grad-CAM visualization is included for qualitative model interpretability. These results demonstrate the synergistic benefit of generative augmentation with self-supervised representation learning for imbalanced medical image classification, with implications for scalable automated DR screening in resource-limited settings.