A novel feature-agnostic approach for glaucoma detection in fundus images

To design and validate a fully automated, feature-agnostic deep learning approach for the early detection of glaucoma from fundus images.

The retrospective study was conducted at Al-Shifa Trust Eye Hospital, Rawalpindi, Pakistan, using publicly available retinal fundus image datasets from October 2024 to January 2025, and comprised 1,707 fundus images from five publicly available retrospective datasets. A computer-aided detection system was employed based on state-of-the-art deep convolutional neural networks, including EfficientNetV2b0, Xception, InceptionV3, Visual Geometry Group and ResNet50. The images were labelled either as glaucomatous or healthy after they were pre-processed through cropping, normalisation and data augmentation. The models were fine-tuned using transfer learning, and evaluated using standard metrics, such as accuracy, precision, recall, F1-score and area under the curve, which were calculated using Python-based statistical libraries.

Among the tested models, EfficientNetV2b0 achieved the best performance with an area under the curve of 0.98 and an accuracy of 93%. The best-performing model achieved a sensitivity/recall of 97%, precision of 89% and F1-score of 93%, indicating reliable performance for glaucoma classification. The robustness of the proposed method was validated across multiple datasets, ensuring its generalisability in diverse clinical scenarios.

The proposed deep learning-based approach provided a reliable and efficient method for early glaucoma detection. Its automation and high accuracy made it suitable for use in mass screening programmes and under-resourced clinical environments, potentially reducing the burden on ophthalmologists and enabling timely intervention.