Deep Learning-based Model Benchmarking of Glaucoma Segmentation Using a Novel Ibn Al-Haitham Fundus Image Dataset

abstract-Automated fundus image analysis is critical in diagnosing glaucoma, whereby the optic nerve areas that consist of optic discs and cups need to be precisely mapped, which is a process that requires considerable effort and time. Generally, the application of Artificial intelligence (AI) has made considerable advancements in the fundus imaging-based automated diagnosis of various diseases. Training and assessing the best AI models for glaucoma detection demands a large number of expert-annotated fundus images. However, there is a scarcity of freely available datasets tailored to this need, and even those that are accessible have several constraints, especially for the Middle Eastern country cases. This paper introduces a novel fundus image dataset of 3000 high-resolution, manually annotated color fundus images dedicated to boosting AI-based models for optic disc and cup segmentation. The annotation process was standardized through collaboration among medical experts, ensuring consistent and reliable results. The reliability of each image was assessed using four deep learning-based segmentation models: U-Net, SegNet, FCDenseNet, and PSPNet. The SegNet model outperformed the others, with an accuracy of 98.12%, specificity of 96.35%, and sensitivity of 98.7% for the optic disc and an accuracy of 97%, specificity of 95.4%, and sensitivity of 91.94% for the optic cup. These results not only validate the accuracy of the developed dataset but also demonstrate its potential in advancing disease detection methods, particularly in AI-assisted ophthalmology, leading to new opportunities for advancements in this field.