<p>Diabetic Retinopathy (DR) is a serious condition affecting diabetic patients, where retinal blood vessels are damaged due to high blood sugar, often leading to blindness. Early detection using computer vision can prevent vision loss. This work introduces a novel Hybrid CST-FWN² model, which combines the Consecutive Swin Transformer (CST) for hierarchical local–global feature extraction with the Fuzzy Wavelet Neural Network (FWNN) for multiscale frequency-domain analysis. Their integration enhances pattern recognition, reduces classification errors, and improves uncertainty handling in medical images. Model parameters are fine-tuned with the proposed Commutable Secretary Bird Optimization (CSBO) algorithm, which outperforms traditional optimizers like Genetic Algorithm and Particle Swarm Optimization in speed and robustness. The pipeline includes preprocessing steps such as noise removal, resizing, and green channel extraction, followed by segmentation using a Hybrid Watershed Technique (HWT) that integrates thresholding and watershed methods for better boundary detection. The extracted features are fused to classify DR into Non-DR, Moderate DR, and Proliferative DR with superior accuracy. Evaluation on the Kaggle EyePACS dataset demonstrates the effectiveness of the approach, achieving 98% accuracy, 95.14% precision, 95.76% recall, 96.43% sensitivity, and 95.03% F1-score, outperforming existing models and offering a scalable automated solution to assist ophthalmologists in DR detection.</p>