Abstract
This study focuses on the development and evaluation of methods for preprocessing and analyzing fundus images to detect diabetic retinopathy (DR), a widespread eye disease in diabetic patients. The main problem addressed is the improvement of image classification accuracy for early DR diagnosis using deep learning techniques. The research utilizes the EfficientNetB0 neural network architecture, with preprocessing steps including image augmentation (rotation, scaling, cropping, contrast enhancement) and normalization to increase model robustness. Two training strategies were tested: (1) freezing the base layers and training only the classification head, and (2) fine-tuning by unfreezing the top layers after initial training. Experimental results showed that fine-tuning significantly improved performance: precision on the test set increased from 65% to 75%, recall from 60% to 74%, and F1-score from 62% to 74%, while accuracy remained stable at 80%. These improvements are explained by the model’s enhanced ability to generalize after fine-tuning, reducing overfitting and capturing more complex features. A distinctive aspect of the results is the comparative analysis of EfficientNetB0 adaptation strategies, which allowed for the identification of the most effective approach. The proposed method demonstrates high potential for implementation in automated diagnostic systems for early detection of retinal diseases and can be applied in clinical decision-support tools under limited data conditions.