Attention U-Net for Early Detection of Diabetic Retinopathy and Glaucoma via Retinal Fundus Image Segmentation

Abstract

Diabetic retinopathy (DR) and glaucoma are two of the most common causes of blindness globally, affecting millions and creating a public health burden. Early identification of both DR and glaucoma is key to avoiding irreversible loss of vision. Conventional diagnosis of DR or glaucoma depends on manual examination of the retinal fundus images by trained professionals. This medical process is laborious and may overlook early subtle changes. To overcome these barriers, this study seeks to investigate an automated deep learning approach with Attention U-Net model for the segmentation and detection of retinal abnormalities related to DR and glaucoma based on retinal fundus data. The Attention U-Net framework accurately segments substantial retinal structures, including blood vessels and the optic nerve head, and places an emphasis on pathological features in these structures such as the microaneurysms related to DR, and damage to the optic nerve related to glaucoma. The model is trained with specialized loss functions, Dice Loss and Focal Loss, in order to mitigate class imbalance and improve sensitivity in detecting the lesions. The model has been trained and validated on public datasets, including DRIVE, DIARETDB1, and RIM-ONE, showing robust and reliable performance. The experimental findings reveal that Attention U-Net achieves better performance than standard segmentation networks, such as U-Net, SegNet, and DeepLab, based on quantitative measures including accuracy, Dice coefficient, intersection over union (IoU), sensitivity, and specificity. Visualizations of the segmentation results indicate that the model demonstrates an improved ability to delineate the complex vascular structure, as well as the optic nerve structure that indicates a potential risk for early diagnosis. In summary, the Attention U-Net framework provides a quick and accurate automated process for ophthalmologists to analyze retinal fundus images for early detection of DR and glaucoma, prevention for timely treatments, and possibly alleviating global impact on vision impairment.