Application of vision transformers in the early detection of excavation in the BRSET base

Enlarged excavation of the optic papilla, caused by the loss of fibres that originate in the retina and transmit electrical stimuli to the visual cortex, is a critical indicator in the early detection of glaucoma, a disease that can lead to irreversible blindness. As the optic papilla shows morphological variations in the population, its identification can be a challenge. Methods based on deep learning have shown promise in helping doctors analyse these images more accurately. Recently, models such as Vision Transformers (ViT) have shown significant results in various medical applications, including glaucoma detection. However, the scarcity of quality data remains a major obstacle to training these models. This study evaluated the performance of the Swin Transformer, DeiT and Linformer models in detecting optic papilla excavation, using the new Brazilian Multilabel Ophthalmological Dataset (BRSET). The results showed that the DeiT model obtained the best accuracy, with 0.94, followed by the Swin Transformer, with 0.88, and the Linformer, with 0.85. The findings of this study suggest that ViT models can not only significantly improve the detection of glaucomatous papillary excavation, but also strengthen Human-Machine Collaboration, promoting more effective interaction between doctors and automated systems in medical diagnosis.