Abstract
Purpose :
Various population-based clinical studies have reported that retinal arteriolar and venular structural change is associated with systemic cardiovascular and cerebral disease. Therefore, automatic identification of the retinal artery/vein (AV) graph structure with high accuracy and efficiency will benefit the quantitative analysis of retinal vessels required by clinical examinations.
Methods :
Vessel segmentation and skeletonization are first performed to obtain the retinal vascular skeleton. For each intersection points that are connected to 3 or more vessels, a cost function is calculated based on the weighted score of vascular angle, color intensity and vessel width, so as to evaluate the probability of every possible vessel connection relationship at intersection points.
The graph reconstruction procedure is performed as a directed tree growing process from root nodes (nearest vessel to disc) to terminal nodes (peripheral end of tree) following the minimal cost function value. The digraph growing procedure is initiated by adding all the root nodes as starting points and iteratively adding child nodes to the tracked tree, until terminal nodes are reached. Vessels from each individual tree can be easily traced from the corresponding root node with the depth-first-search algorithm. Finally, the AV label for each tree graph is classified using color intensity features with machine learning algorithms and further refined with intertwine relationships among multiple tree graphs.
Results :
The proposed procedure has been tested on the publicly available AV-DRIVE and INSPIRE-AVR databases with pixel level accuracy of 93.03% and 90.78% respectively, which achieves state-of-the-art accuracy for automatic AV classification on the two databases. Fig1 shows the processing results for two images from the databases.
The algorithm takes a total of 35 seconds per image for AV-DRIVE and 37 seconds per image for INSPIRE-AVR database, which is the fastest program compared to the others reported for similar purpose.
Conclusions :
We have developed a tree growing method to identify the retinal AV graph structures. The method is accurate and computationally efficient as tested on two publically available databases.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.