Retinal diseases remain a primary cause of irreversible vision loss worldwide, making early and precise diagnosis of conditions like diabetic retinopathy, glaucoma, and age-related macular degeneration critically important. However, effective retinal image analysis is challenged by factors such as noise, low contrast, and complex anatomical structures, especially when detecting fine vessel boundaries and subtle pathological lesions. To address these limitations, this research proposes a novel Retinal Efficient Deep Metric Learning Framework (RetinaEX-Net) that combines edge-aware image enhancement, high-dimensional embedding, and adaptive classification for robust disease detection. The framework introduces a Smart Adaptive Retinal Pre-processing (SARP) method, which integrates CLAHE (Contrast Limited Adaptive Histogram Equalization) and bilateral filtering, followed by Canny edge detection to enhance structural detail while minimizing artifacts. Feature embeddings are extracted using EfficientNet, with clustering performed using k-Means based on Empirical Bregman Divergence, and classification achieved through a confidence-aware k-NN approach. This architecture enables precise discrimination even in early-stage or ambiguous cases. Experimental evaluations on RFMiD and APTOS 2019 datasets show that RetinaEX-Net achieves up to 98.86% accuracy and a Cohen’s Kappa score of 98.24%, outperforming state-of-the-art models. RetinaEX-Net offers a scalable, interpretable, and clinically deployable artificial intelligence (AI) solution, where AI specifically refers to the integration of deep metric learning, efficient convolutional feature extraction (EfficientNet), adaptive clustering using Bregman divergence, and confidence-aware k-NN classification. This makes the system well-suited for real-time screening in tele ophthalmology and resource-constrained healthcare environments.