Early glaucoma detection using hybrid models based on deep learning in fundus images

Ramírez-Cenepo L.E.; Ruiz-Cueva J.A.; Sánchez-Flores J.; Rojas-Córdova K.L.; Leveau-Paredes M.G.; Llontop-Reategui A.R. · Revista Cientifica de Sistemas e Informatica (2026) · DOI: 10.51252/rcsi.v6i2.1642

Resumen

Glaucoma is one of the leading causes of irreversible blindness, making early detection essential to prevent vision loss. This study aimed to develop and evaluate hybrid deep learning models for multiclass glaucoma classification using fundus images. A total of 6,014 retinal images from public databases were used, after undergoing cleaning, balancing, and stratified splitting. The VGG16, ResNet50, and EfficientNet-B0 architectures were employed as feature extractors combined with Cascaded Deep Forest. Evaluation included stratified 10-fold cross-validation and an independent test set. VGG16 + CDF achieved the best performance, with an average accuracy of 0.9320, greater stability, and better generalization. Statistical tests confirmed significant differences among the models, demonstrating that the proposed hybrid approach is a robust alternative for automated glaucoma classification.

Instituciones: Universidad Nacional de San Martín