Deep Learning Approaches for Diabetic Retinopathy: A Comparative Study Based on Loss Function Optimization

This research aims to study and analyze the performance of loss functions in training deep learning models to detect diabetic retinopathy (DR). The effect of three main loss functions was tested using Freeze and Fine-tuning training techniques with different learning rate scheduling. A dataset of 87,621 fundus images was used, where a model based on convolutional neural networks (CNNs) was trained using transfer learning mechanisms. The model's performance was also compared with previous studies to reach the optimal methodology for building a more accurate classification model. The results showed that choosing the appropriate loss function and optimizing the training parameters lead to a significant improvement in the accuracy of the model, which contributes to enhancing the effectiveness of early detection of diabetic retinopathy using artificial intelligence.