Diabetic retinopathy (DR) is a leading cause of blindness worldwide, particularly affecting individuals with diabetes mellitus. Early detection and treatment are crucial to prevent severe vision loss. Recent advancements in artificial intelligence (AI) and deep learning (DL) have shown promise in automating the detection and classification of DR from retinal fundus images. However, the efficacy of these models is often limited by the quality of the input images, which can vary significantly due to factors such as poor lighting, occlusions, and image artifacts. This study proposes an enhanced image preprocessing pipeline combined with state-of-the-art optimization techniques to improve the performance of AI models for DR detection. The experimental setup involved using three different convolutional neural network (CNN) architectures: ResNet50, EfficientNetB0, and a custom lightweight CNN. These models were trained and evaluated on publicly available datasets, including EyePACS, Messidor, and APTOS 2019 Blindness Detection, which contain high-resolution retinal images labeled according to the severity of DR. The proposed image preprocessing techniques included Contrast-Limited Adaptive Histogram Equalization (CLAHE), Gaussian smoothing, circular cropping, optic disc removal, normalization, and green channel extraction. Additionally, advanced optimization techniques such as data augmentation, Bayesian hyperparameter optimization, dropout, and L2L2L2 regularization were employed to enhance model performance. The results demonstrate that EfficientNetB0 achieved the highest accuracy (94.8%) and AUC (97.8%), indicating its superior ability to differentiate between different levels of DR severity. The model also showed the best balance between sensitivity (93.6%) and specificity (96.1%), making it highly reliable for both detecting true DR cases and correctly identifying non-DR cases. ResNet50 also performed well, with an accuracy of 93.4%, sensitivity of 91.8%, specificity of 95.2%, and an AUC of 96.5%. The custom lightweight CNN, designed for deployment in resource-constrained environments, achieved a slightly lower accuracy of 92.1% and an AUC of 95.2%, but still provided a viable option for real-world applications where computational efficiency is crucial. Overall, the study highlights the importance of combining advanced image preprocessing techniques with robust AI models and optimization strategies to improve DR detection accuracy and reliability. The findings suggest that employing a well-optimized EfficientNetB0 model, alongside comprehensive image preprocessing, offers the best trade-off between accuracy, computational efficiency, and generalizability across diverse clinical settings.

Diabetic retinopathy (DR), Machine Learning, Artificial Intelligence, Transfer Learning, Convolutional Neural Network Algorithm(CNN).