Retinal image segmentation with deep learning models has a great deal of promise to increase ophthalmology diagnostic precision. But these models’ intrinsic complexity frequently makes it difficult to understand and interpret the segmentation results. In past research, there have been various methods proposed for segmentation. However, the challenges still exist using the deep learning methods in medical healthcare. In this study, a novel method to improve the transparency and interpretability of retinal image segmentation is proposed by combining the EfficientNet and Explainable AI (XAI) methodologies with the U-Net architecture. Through the use of Local Interpretable Model-agnostic Explanations (LIME), our methodology produces feature visualizations that clarify the U-Net model’s decision-making process, enabling doctors to confirm and understand the segmentation outcomes. The performance metrics measured are Precision, Recall, and F1-score. The precision obtained by means of LIME is 80%. By bridging the gap between high-performance segmentation models and their useful applications in clinical settings, to improve model generalizability across various datasets, the study uses sophisticated data augmentation techniques such as noise addition, brightness fluctuations, and random rotation. The study also points out drawbacks, such as processing overhead and the effect of dataset diversity and size on model performance, paving the potential for further improvements. This integration of XAI approach seeks to encourage trust and enable more informed medical decisions.