Retinal disorders such as diabetic retinopathy, glaucoma, and macular degeneration are among the leading causes of global visual impairment, affecting over 2.2 billion people worldwide, many of whom could benefit from early diagnosis. However, the development of reliable AI-based retinal screening systems is hindered by severe class imbalance in clinical datasets, where common diseases dominate, and rare conditions are poorly represented. This study proposes an Imbalance-Resilient Dual-Balanced Transfer Learning Framework to address this challenge. The framework combines weighted random sampling at the data level with class-balanced focal loss at the optimization level, built upon a fine-tuned EfficientNet-B0 backbone for multiclass retinal disease classification. Weighted sampling enhances minority-class exposure without introducing synthetic artifacts, while focal loss (γ=2) emphasizes hard-to-classify samples, enabling improved detection of subtle retinal abnormalities. The proposed method was evaluated on 21,577 fundus images spanning ten retinal disease categories. Experimental results demonstrate an average accuracy of 99.84%, a macro-F1 score of 97.59%, and robust performance on rare conditions, achieving an F1-score of 83.87% for pterygium. Overall, the framework provides an efficient, fair, and generalizable solution for imbalanced retinal disease classification, supporting the development of clinically viable and bias-resistant ophthalmic AI systems.