Diabetic retinopathy (DR) is a leading cause of vision loss among individuals with diabetes. Early detection and precise classification of DR stages are essential for timely medical intervention. This study introduces a novel approach for DR detection utilizing the NASNetLarge model on the APTOS 2019 dataset. To address class imbalances, extensive data augmentation techniques and a balanced class weighting strategy were employed. The NASNetLarge model, trained on the DR dataset, achieved an accuracy of 97%, demonstrating its efficacy in identifying DR stages. Early stopping and learning rate reduction were applied to prevent overfitting and enhance generalization. The proposed method not only enhances detection accuracy but also maintains high performance across various DR stages, presenting a promising tool for automated DR screening and early diagnosis.