Retinal Arteries and Veins Semantic Segmentation Using Modified U-Net Architecture

This paper presents a modified U-Net architecture designed for the semantic segmentation of retinal arteries and veins using infrared reflectance retinal imaging. Leveraging the RAVIR dataset, which comprises 46 high-quality retinal images, we focus on addressing the challenges of low contrast and varying lighting conditions often found in retinal images. Our approach incorporates advanced features such as bilinear interpolation and attention mechanisms to enhance segmentation accuracy. The modified U-Net architecture achieves a mean Intersection over Union (mIoU) of 0.8855 and a Dice coefficient of 0.9382, showing a significant improvement over the traditional U-Net. These results demonstrate its potential for practical applications in retinal disease diagnosis and monitoring.