Diabetic retinopathy (DR) is a major cause of vision impairment and blindness worldwide, necessitating precise and early diagnosis to enable timely intervention. This study presents an optimized framework for DR detection and classification, integrating advanced segmentation and hybrid feature extraction techniques to address challenges in robustness, scalability, and interpretability. The methodology employs a Firefly heuristic-driven Fuzzy C-Means (FAFCM) algorithm for precise segmentation of pathological features, including microaneurysms and hemorrhages. Spatio-textural features extracted via the Gray-Level Co-occurrence Matrix (GLCM) are combined with deep features from ResNet101 and DenseNet121 to capture structural and contextual information. These features undergo dimensionality reduction and are classified using a robust ensemble of SVM, DT, and RF, consolidated through maximum voting. Evaluated on EyePACS and Messidor-2 datasets, the framework achieves 99.2% accuracy, demonstrating better generalization over existing techniques, and establishing a potential for clinical applications in DR diagnosis.