Enhanced real-time glaucoma diagnosis: dual deep learning approach

Abstract

Effective management of glaucoma is essential for preventing irreversible vision loss. This study introduces a novel deep learning-based network designed to enhance performance while minimizing computational complexity. The system comprises two models: the first is a hybrid model combining a customized U-Net architecture integrated with you only look at coefficients (YOLACT) is utilized to achieve accurate segmentation of the optic disc (OD) and optic cup (OC), providing detailed diagnostic insights for ophthalmologists. The second model employs you only look once version 5 (YOLOv5) for real-time glaucoma prediction, delivering outstanding performance with an accuracy of 97.89% and F1 score of 98% on the primary dataset. On an independent dataset without further training, the model achieved 96% accuracy, with sensitivity and specificity of 98.9% and 93.3%, respectively. These results highlight the model's robustness, generalizability, and adaptability, demonstrating its potential for effective glaucoma screening and early detection in diverse clinical environments. This approach offers a promising advancement in improving the accessibility and efficiency of glaucoma management.