Globally, over 2.2 billion people suffer from vision impairment due to retinal diseases, such as diabetic retinopathy, glaucoma, and cataracts. Early and accurate detection is critical, yet existing deep learning models face challenges such as inadequate interpretability, suboptimal feature extraction, and a lack of trust from medical professionals, which hinders their clinical adoption. To bridge this gap, our proposed work modified the dense net with Grad cam++ to provide both high diagnostic performance and improved transparency, ensuring trust, and validating AI-generated results. We trained and evaluated our model using a diverse retinal image dataset covering three major diseases: diabetic retinopathy, glaucoma, and cataracts. The model achieved an impressive accuracy of 94.88%, with a precision of 94.67% and recall of 94.56%, significantly outperforming conventional deep-learning models. The integration of Grad-CAM++ further enhanced interpretability by generating heatmaps that accurately highlighted disease-specific regions closely aligned with expert annotations. The proposed model ensures reliable diagnostics while bridging the gap between AI predictions and clinical decision-making. Real-time interpretable results aid early detection and improve patient outcomes.