Enhanced Diabetic Retinopathy Identification via EfficientNetB3-Based Convolutional Neural Networks and Transfer Learning

Diabetic retinopathy (DR) is one of the most common causes of blindness worldwide, making early detection essential for treating this disease. This research presents the creation and testing of a convolutional neural network (CNN) model based on the EfficientNetB3 architecture to detect signs of DR in typical fundus images. The model was trained with a set of retinal images and tested using metrics such as accuracy, recall, and F1 score. The results show a weighted accuracy of 81%, with high performance in the healthy class (97% accuracy and 98% recall). However, lower accuracy is observed in the advanced stages of DR, mainly attributed to class imbalance in the dataset. It is also necessary to balance the classes and combine the architecture with other models. These findings demonstrate the potential of EfficientNetB3 as a diagnostic support tool and highlight the need to improve data balance and the model for better discrimination of severe cases.