Deep learning framework for glaucoma detection using enhanced U-Net segmentation and 2D-CNN classification

Abstract

Glaucoma is a major global health concern and the foremost cause of irreversible blindness, second only to cataracts in prevalence. Early diagnosis is challenging since the disease progresses slowly with minimal symptoms. Ophthalmologists typically evaluate optic nerve damage, often relying on the Disc Damage Likelihood Score (DDLS). Another critical indicator is the retinal nerve fiber layer (RNFL), which thins around the optic nerve head (ONH) in glaucomatous eyes. Traditional manual assessments, however, demand expertise and time, limiting their practicality. The study developed a deep learning-based method for glaucoma sickness identification using RNFL features. The technique ensured accurate retinal structure extraction by segmenting using an improved U-Net model. We used data augmentation approaches to enhance the generalization of the model. Many examples were used to train the model, which was then fine-tuned for accuracy and boundary preservation. To distinguish between glaucomatous and non-glaucomatous cases based on anatomical variations, the extracted features were analyzed using a two-dimensional convolutional neural network, which served as the primary tool for classification. The suggested technique effectively separated the RNFL from OCT pictures, guaranteeing precise extraction of structural information essential for glaucoma classification, and it obtained a classification accuracy of 96.1% on a retinal imaging dataset.