From Laboratory to SUS: Addressing Data Selection, Model Uncertainty, and Realism in Diabetic Retinopathy Lesion Detection Using a Brazilian Fundus Dataset

• Laura Bernardes IFFar
• Alejandro Pereira UFPel
• Marcelo Dias UFPel
• Marilton Aguiar UFPel
• Daniel Welfer UFSM
• Carlos Santos IFFar

DOI: https://doi.org/10.5753/sbsi_estendido.2026.249073

Resumo

Diabetic retinopathy (DR) is a leading cause of vision loss in working-age adults. Early diagnosis depends on ophthalmological examination, but access remains limited in Brazil’s public healthcare system (Sistema Único de Saúde — SUS) due to infrastructure and specialist shortages. Deep learning can support screening and triage, but progress is constrained by the scarcity of publicly labeled data. We curated and anonymized 13,131 color fundus images collected between 2012 and 2024 and selected 150 images to create an initial Brazilian dataset with DR grading and lesion annotations. We evaluated representative deep neural networks for classification, semantic segmentation, and lesion detection. RegNet Y 400MF achieved the best grading results (overall accuracy 0.6667, average accuracy 0.4634, Cohen’s kappa 0.5179), R2U-Net showed the most balanced segmentation performance, and YOLOv11 obtained the highest mAP (0.3460 on validation; 0.3810 on test). The results highlight practical challenges, especially for microlesions, and support a SUS-oriented, deployment-aware perspective with attention to uncertainty, referral decisions, and cross-site robustness.

Referências

Aghelan, A. and Rouhani, M. (2022). Fine-tuned generative adversarial network-based model for medical image super-resolution. arXiv preprint arXiv:2211.00577.

Chakrabarti, R., Harper, C. A., and Keeffe, J. E. (2012). Diabetic retinopathy management guidelines. Expert Review of Ophthalmology, 7(5):417–439.

Li, T., Gao, Y., Wang, K., Guo, S., Liu, H., and Kang, H. (2019). Diagnostic assessment of deep learning algorithms for diabetic retinopathy screening. Information Sciences, 501:511–522.

Nandal, P., Pahal, S., Khanna, A., and Pinheiro, P. R. (2024). Super-resolution of medical images using real esrgan. IEEE Access.

Nayak, J., Bhat, P. S., Acharya U, R., Lim, C. M., and Kagathi, M. (2008). Automated identification of diabetic retinopathy stages using digital fundus images. Journal of Medical Systems, 32(2):107–115.

Pereira, A., Santos, C., Aguiar, M., Welfer, D., Dias, M., de Menezes, R., and Santana, D. (2024). Enhanced semantic segmentation of retinal microlesions through r2u-net architecture. In Simpósio Brasileiro de Computação Aplicada à Saúde (SBCAS), pages 13–24, Goiânia, GO. SBC.

Porwal, P., Pachade, S., Kokare, M., Deshmukh, G., Son, J., Bae, W., Liu, L., Wang, J., and Liu, X. (2020). IDRiD: Diabetic Retinopathy – Segmentation and Grading Challenge. Medical Image Analysis, 59.

Rashid, S. I., Shakibapour, E., and Ebrahimi, M. (2022). Single mr image super-resolution using generative adversarial network.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net: Convolutional networks for biomedical image segmentation. Lecture Notes in Computer Science (including sub-series Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 9351:234–241.

Vocaturo, E. and Zumpano, E. (2020). The contribution of AI in the detection of the Diabetic Retinopathy. In Proceedings of the IEEE International Conference on Bioin formatics and Biomedicine, BIBM 2020, pages 1516–1519, Seoul, Korea, 16–19 December 2020. IEEE.

Wang, X., Xie, L., Dong, C., and Shan, Y. (2021). Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In International Conference on Computer Vision Workshops (ICCVW).