Edge-Aware Vessel Segmentation Based on Local-to-Global Supervision

‌DRIVE Dataset‌: The DRIVE dataset consists of 40 fundus images collected from 7 patients with retinal disease and 33 healthy individuals. Each image has a resolution of 584 × 565 pixels. A total of 20 images are used for training, while the remaining 20 images are designated for testing. Each image has two annotations provided by two different experts, and we use the first annotation as the ground truth.‌CHASE_DB1 Dataset‌: The CHASE_DB1 dataset contains 28 retinal images captured from the left and right eyes of 14 children. Each image has a resolution of 999 × 960 pixels. The first 20 images are used for training, while the remaining images serve as test samples. Similar to the DRIVE dataset, each image is annotated by two experts, and we use the first annotation as the ground truth.‌DCA1 Dataset‌: The DCA1 dataset consists of 134 X-ray coronary angiography images with ground truth annotations provided by cardiology experts. The entire database is sourced from the Mexican Social Security Institute, UMAE T1-Leon. Each angiographic image is in PGM format with a grayscale resolution of 300 × 300 pixels. The dataset is divided into two subsets: 100 images for training and the remaining 34 images for testing.