Optimization of retinal artery/vein classification based on vascular topology

Coronary heart disease (CHD) poses a serious threat to human health. The retinal vessel research is an important direction for the non-invasive diagnosis of CHD. Convolutional neural network (CNN) used for artery and vein (A/V) classification has a main problem such as the A/V misclassification, that severely affects the A/V parameters extraction and CHD diagnosis. In this paper, a post-processing optimization method is proposed to solve the problem to improve the accuracy of A/V classification. A backflow tracing algorithm is proposed to get the vascular directed topology, which tracks starting from the end of the vessel to the optic nerve head (ONH) to avoid repeating traversal of trunk vessels. Considering the influence of vascular topology structure on the vascular label, an A/V label delivery algorithm is proposed to obtain the directed vessel with A/V labels based on the types of branch nodes. Experiments clearly show that the post-processing optimization method performs well on three datasets with different scales. By utilizing the vascular topology, the post-processing optimization method greatly improves the A/V classification accuracy of the results from CNN. The proposed method improves the accuracy of A/V classification. The optimization results can be used to extract vascular parameters, such as tortuosity and caliber, which are expected to be used in the intelligent diagnosis of CHD.