Artificial Intelligence-Based Screening for Diabetic Retinopathy: Model Comparison and Interpretability

Diabetic Retinopathy is one of the common complications of diabetes and can lead to permanent vision loss if left untreated. This study examined the performance of different AI-based methods for DR classification. Deep learning-based models, ResNet-50, DenseNet-121, U-Net, and classical CNN structures, along with traditional machine learning algorithms, SVM, Decision Trees, and k-Nearest Neighbor, were evaluated on the APTOS 2019 dataset. To optimize model performance, image data were subjected to various preprocessing steps, such as resizing, contrast correction, and denoising. Augmentation techniques were used to increase data diversity. According to experimental results, the most successful model was DenseNet-121, with an accuracy rate of 87% and an F1 score of 86%. In contrast, while classical machine learning methods produce lower accuracy values than deep learning, they exhibit consistent performance under certain conditions and offer a more computationally cost-effective alternative. The comparisons indicate the applicability of classical methods, especially in scenarios with limited data. This evaluation process creates a basic framework that will enable the integration of explainable artificial intelligence (XAI) approaches in later stages and is a preparation for adapting interpretation techniques such as SHAP and LIME to clinical decision support systems.