Transformer-based models significantly enhance diabetic retinopathy detection and classification compared to traditional methods. However, challenges remain, particularly the need for large labeled datasets and the difficulty of interpreting model results. If integrated into clinical practice, these models could transform diabetic retinopathy diagnosis, enabling earlier, more precise interventions, reducing diagnostic costs, and increasing access to quality visual health care. Overall, while promising, further research is needed to optimize these models for clinical use, particularly in balancing performance with computational demands. A systematic review of studies from 2018 to 2024, following PRISMA guidelines, was conducted to assess the accuracy of these models, efficiency, and clinical relevance. The main findings of this research are: the integration of advanced models such as transformers could facilitate clinical decision-making, optimizing the time and resources of healthcare professionals, but while transformers can offer accurate predictions, they often fail to provide sufficient insight into how those predictions were reached, making clinical adoption difficult, this opens an opportunity for the development of intelligent computing models with capabilities to show the how the output of the algorithm is generated.