Enhancing Glaucoma Detection with Deep Learning: Segmentation of Optic Nerve Features from Low Quality Images

Glaucoma is a leading cause of irreversible blindness worldwide, with early detection critical for effective intervention. However, access to reliable screening remains limited in many under-resourced settings due to reliance on high-quality imaging and specialist interpretation. This study introduces a deep learning-based system built around an enhanced U-Net architecture for segmenting optic nerve head features from both high-resolution fundus photographs and lowquality clinical video ophthalmoscopy frames. The model uses an encoder-decoder architecture to improve robustness under challenging imaging conditions, such as glare and motion blur. Trained on the ORIGA dataset and evaluated on real-world clinical images collected from high-risk patients, the U-Net model demonstrates strong segmentation performance, particularly when optimized with a hybrid loss function combining binary cross-entropy and Dice loss. Comparative evaluation with a promptable Medical SAM-2 model highlights the UNet’s superior reliability in maintaining accurate optic disc and cup boundaries, especially for the more subtle optic cup region. This work underscores the potential of lightweight, interpretable U-Net-based models to support scalable, equitable glaucoma screening in diverse clinical environments, laying the groundwork for future real-time deployment and integration with diagnostic scoring systems like DDLS.