Extraction of Network of Blood Vessels in Fundus Images

Abstract

Proposed algorithm tracks network of blood vessels and evaluates diameter of vessel in each point of extracted network. Algorithm utilizes opposite-parallel vessel tracking method (tracking the points of opposite blood vessel edges). Introduced novelty in the algorithm: edges of blood vessels were found using also gradient vector phase matrix, calculated for entire image. In addition, quantitative measures were introduced in order to evaluate the performance of the algorithm to detect vascular network. The first measure shows percentage of correctly detected points by the algorithm. If we subtract the value of this measure from 100 % we receive an error rate. The second measure shows percentage of automatically detected vascular network. This measure could be used as detection rate. These two measures correspond to the strongest condition: the detected points by the algorithm must coincide with the manually labeled vascular network in the real test images (8 images were used). Other two measures were obtained when coincidence condition was relaxed: the detected points must refer into either exactly to labeled points or to neighboring points (±1 pixel) of test images. By using these measures we optimized the algorithm and received the following results: 66,7 % average detection rate and 11,21 % average error rate, with relaxed conditions. Ill. 6, bibl. 6 (in Lithuanian; summaries in Lithuanian, English and Russian).