Abstract

Accurate segmentation of retinal blood vessels and automated detection of abnormalities play a crucial role in the early diagnosis of ophthalmic and systemic diseases. However, low-contrast retinal images significantly hamper the performance of traditional segmentation methods. We formulated a hybrid deep learning model that combines three advanced deep learning architectures, such as GANs for image enhancement, Transformer encoders for global contextual representation, and a Net for fine-grained vessel segmentation. The GAN component enhances the image quality by increasing the contrast and structural details. On the other hand, the Transformer encoder is responsible for capturing the long-range dependencies across retinal areas. Finally, the Attention U, Networks on further segment refinement by focusing on the vascular structures that are clinically relevant. Such a synergistic combination allows the model to be very robust on challenging datasets, here achieving a segmentation accuracy of 99%. Besides vessel extraction, the framework also supports automatic abnormality deduction, thus allowing early detection of pathological changes. Experiments show that the proposed hybrid model distances significantly from the existing ones, especially in the case of low signal-to-noise, hence pointing to its potential as a trustworthy tool for computer-aided retinal disease screening and diagnosis.