Early detection of diabetic retinopathy (DR), particularly in its earliest stage marked by microaneurysms, is crucial for preventing vision loss. Our study integrates advanced AI technologies to develop a more efficient and transparent system for diagnosing DR. By leveraging Capsule Networks (CapsNets) and Explainable AI (XAI), we aim to enhance both the accuracy of identifying retinal lesions and the interpretability of the diagnostic process. The proposed CapsNet model achieves a 94% accuracy, representing a 9% improvement over the baseline CNN model. Additionally, it achieves a 5% improvement in F1-score, precision, and recall, demonstrating its effectiveness in small lesion detection. Our use of Grad-CAM, an XAI technique, provides visual insights into the regions of retinal images that influence the model’s predictions, fostering trust and validation in clinical settings. By combining improved accuracy with interpretability, our approach offers a powerful tool to streamline DR screening, enabling clinicians to make quicker and more reliable diagnoses without extensive manual efforts.