Building Robust binary Pre-trained Deep Learning Model For Diabetic Retinopathy Detection Using Local Datasets

One of the main challenges in automated Diabetic Retinopathy (DR) detection is the heterogeneity of available datasets. Models trained on one dataset often experience a sharp drop in performance when evaluated on images from other sources, due to variations in imaging devices, acquisition protocols, patient demographics, and annotation standards. This domain shift limits the generalizability of DR screening systems and highlights the need for robust pretrained models that can adapt to diverse data distributions. In this study, we developed a binary classification framework for DR screening using a pretrained ResNet50 convolutional neural network. We evaluated four experimental approaches: (1) training the model solely on a high-quality APTOS dataset, (2) training on the large-scale EyePACS dataset, (3) merging public datasets for more diverse representation, and (4) applying transfer learning and selective fine-tuning to adapt the model to new data domains. Results show that a model trained solely on the high-quality APTOS dataset achieves the highest accuracy (97%), while merging heterogeneous datasets improves robustness across external datasets despite a trade-off in positive case recall. Transfer learning with selective fine-tuning further enhances adaptability to new domains, reaching 87% accuracy on unseen datasets. Moreover, our customized pretrained model, after training on merged datasets, consistently outperforms the baseline ResNet50 pretrained on ImageNet, demonstrating the advantage of domain-specific adaptation. These findings provide actionable insights for developing robust DR classification models with strong generalization potential for clinical applications.

• Diabetic Retinopathy
• Fine-Tuning
• Model Adaptation
• Pre-Trained Model
• Transfer Learning