Masked Vascular Structure Segmentation and Completion in Retinal Images

This paper presents a U-Net–based deep learning framework for accurate retinal blood vessel segmentation from grayscale fundus images. The proposed model employs a two-class semantic segmentation approach distinguishing vessel and background pixels. The dataset is prepared using image and pixel label datastores, where images and ground truth masks are resized, normalized, and validated for consistency. Data augmentation techniques such as rotation, reflection, and translation improve model robustness against variations in orientation and illumination. A customized U-Net architecture with class weighting is used to handle pixel imbalance, ensuring enhanced vessel detection performance. The network is trained using the Adam optimizer with a low learning rate to achieve stable convergence. After prediction, softmax activation maps and thresholding refine vessel probability estimation. Post-processing operations like morphological closing and hole filling eliminate false detections. Quantitative evaluation using metrics such as Dice, Jaccard, clDice, CAL, and connectivity loss (LC) demonstrates strong structural consistency and accurate vessel delineation compared to ground truth.

Keywords: Retinal vessel segmentation, U-Net, semantic segmentation, data augmentation, class weighting, softmax activation, post-processing, connectivity loss, Dice coefficient, Jaccard index.