Multi-stage generative adversarial network model for segmenting retinal vascular structures in eye disease prediction.

Abstract

Retinal vessel segmentation is essential for precise ophthalmological diagnoses, particularly in the prediction of retinal degenerative diseases. However, existing methods usually lack robustness and accuracy, especially in segmentation of thin or overlapping vessels. To face these challenges, this study introduces an enhanced retina-RV-Gain segmentation model, which employs an architecture of various stages to refine the results of segmentation iteratively. The model integrates attention mechanisms to better capture complex vessel structures and employs an adaptive loss function to manage class imbalance. In addition, a specially designed discriminator enhances the model's ability to distinguish fine details from background noise vessels. The proposed RV-Gan is trained in comprehensive data sets that comprise retinal images, segmentation masks and noted labels, including Stare-DB, Chase-DB1 and Drive, using the Python platform. Experimental results demonstrate a segmentation accuracy of up to 99% in normal, abnormal and base vessels. These findings highlight the potential of the model to significantly improve diagnostic accuracy and support early prediction of disease in clinical ophthalmology. Overall, the enhanced RV-Gan architecture offers a robust solution to the limitations of current approaches, providing segmentation of high fidelity retinal vessels and advancing the predictive analysis of retinal degenerative conditions.