Improving Diabetic Retinopathy Lesion Detection and Segmentation via a Macula-Guided Region of Interest Pipeline Based on YOLOv12
Resumo
Diabetic retinopathy is a leading cause of vision impairment worldwide and is diagnosed through retinal lesions such as hard exudates, hemorrhages, soft exudates, and microaneurysms. Among these, microaneurysms represent the earliest clinical manifestation but remain difficult to detect due to their small size and low contrast. This study proposes a macula-guided Region of Interest (ROI) pipeline to improve retinal lesion detection and segmentation by exploiting the anatomical relevance of the macular region, where lesions are often more concentrated. The proposed framework employs two YOLOv12-based models: the first detects the optic disc and macula to extract ROIs, while the second detects and segments lesions using these regions during training. Experiments conducted on the Dataset for Diabetic Retinopathy achieved an mAP@50 of 0.3240 for detection and 0.3340 for segmentation. To assess generalization, the model was further evaluated on the Indian Diabetic Retinopathy Image Dataset, obtaining an mAP@50 of 0.3990 for detection and 0.2910 for segmentation. The results indicate that anatomically guided region-based training may improve predictive performance, particularly for subtle and small-scale lesions.
Referências
Dias, M., Santos, C., Pereira, A., Aguiar, M., and Welfer, D. (2025). Enhanced detection and segmentation of diabetic retinopathy lesions using a region of interest strategy. In Anais do LII Seminário Integrado de Software e Hardware, pages 61–72, Porto Alegre, RS, Brasil. SBC.
Dias, M., Santos, C., Pereira, A., Aguiar, M., and Welfer, D. (2026). Retinal lesion detection and segmentation using a region of interest-based approach with yolo11. In de Freitas, R. and Furtado, D., editors, Intelligent Systems, pages 440–454, Cham. Springer Nature Switzerland.
Elloumi, Y. and Kachouri, R. (2023). A robustness study of machine learning based methods for macula detection in pathological fundus images. In 2023 Twelfth International Conference on Image Processing Theory, Tools and Applications (IPTA), pages 1–6.
International Council of Ophthalmology (2017). Updated 2017 ico guidelines for diabetic eye care. ICO Guidelines for Diabetic Eye Care, pages 1–33.
Jocher, G., Chaurasia, A., and Qiu, J. (2023). Ultralytics yolov8.
Jocher, G. and Qiu, J. (2024). Ultralytics yolo11.
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.
Martins, J., Coelho, A., and Baffa, M. (2021). Detecção dos estágios iniciais da retinopatia diabética utilizando redes neurais profundas. In Proc. of the VIII Escola Regional de Computação Aplicada à Saúde (ERCAS 2021), pages 42–45, Porto Alegre, RS, Brasil. SBC.
Munuera-Gifre, E., Saez, M., Juvinyà-Canals, D., Rodríguez-Poncelas, A., Barrot-de la–Puente, J.-F., Franch-Nadal, J., Romero-Aroca, P., Barceló, M. A., and Coll-de Tuero, G. (2020). Analysis of the location of retinal lesions in central retinographies of patients with type 2 diabetes. Acta Ophthalmologica, 98(1):e13–e21.
Ometto, G., Assheton, P., Calivá, F., Chudzik, P., Al-diri, B., Hunter, A., and Bek, T. (2017). Spatial distribution of early red lesions is a risk factor for development of vision-threatening diabetic retinopathy. Diabetologia, 60(12):2361–2367.
Pereira, A., Santos, C., Aguiar, M., Welfer, D., Dias, M., and Ribeiro, M. (2023). Improved detection of fundus lesions using yolor-csp architecture and slicing aided hyper inference. IEEE Latin America Transactions, 21(7):806–813.
Porwal, P., Pachade, S., Kamble, R., Kokare, M., Deshmukh, G., Sahasrabuddhe, V., and Meriaudeau, F. (2018). Indian diabetic retinopathy image dataset (idrid): A database for diabetic retinopathy screening research. Data, 3(3).
Romero-Aroca, P., Garcia-Curto, E., Pascual-Fontanilles, J., Valls, A., Moreno, A., and Baget-Bernaldiz, M. (2024). Distribution of microaneurysms and hemorrhages in accordance with the grading of diabetic retinopathy in type diabetes patients. Diagnostics, 14(14).
Santos, C., Aguiar, M., Welfer, D., Dias, M., Pereira, A., Ribeiro, M., and Belloni, B. (2023). A new approach for fundus lesions instance segmentation based on mask r-cnn x101-fpn pre-trained architecture. IEEE Access, 11:43603–43618.
Tian, Y., Ye, Q., and Doermann, D. (2025). Yolov12: Attention-centric real-time object detectors.
Wang, C.-Y., Yeh, I.-H., and Liao, H.-Y. M. (2024). Yolov9: Learning what you want to learn using programmable gradient information.
