| Abstract [eng] |
This paper aims to compare different modern eye disease classification methods using deep neural networks. To carry out the research, convolutional neural networks and vision transformers were analyzed, their architectures used in the field of medical image processing were discussed. Training methods were also discussed, with supervised training and transfer learning being the most commonly used for classification tasks. The APTOS 2019 dataset was chosen for the work and data processing and preparation methods were selected based on the literature. During the study, 8 experiments were performed with different configurations, three metrics and the error were used and compared for more accurate evaluation. The benchmark results showed that vision transformers can achieve marginally but better results than the selected EfficientNet network, while vision transformers trained using transfer training can achieve significantly better results but may overfit. The conclusions of the work contain suggestions on what further experiments should be carried out and what should be explored in more detail in order to better evaluate the models, achieve better results and prepare the models for application outside of research. |
