An improved retinal vessel segmentation method based on high level features for pathological images

Most of the retinal blood vessel segmentation approaches use low level features, resulting in segmenting non-vessel structures together with vessel structures in pathological retinal images. In this paper, a new segmentation method based on high level features is proposed which can process the structure of vessel and non-vessel independently. In this method, segmentation is done in two steps. First, using low level features segmentation is accomplished. Second, using high level features, the non-vessel components are removed. For evaluation, STARE database is used which is publicly available in this field. The results show that the proposed method has 0.9536 accuracy and 0.0191 false positive average on all images of the database and 0.9542 accuracy and 0.0236 false positive average on pathological images. Therefore, the proposed approach shows acceptable accuracy on all images compared to other state of the art methods, and the least false positive average on pathological images.