The Arteriovenous Classification in Retinal Images by U-net and Tracking Algorithm

Abstract

Retinal vessel is the only vessel structure in human circulatory system that can be directly observed by non-invasive methods. According to clinical findings, the reduction of arteriovenous width ratio (AVR) acts as an indicator to predict the risk of many systemic diseases. Therefore, it's essential to develop an automatic classification method for arteries and veins to calculate AVR. A method that combines the deep segmentation network and tracking algorithm is proposed in this paper to classify arteries and veins in retinal images. This automatic processing has three steps: (1) retinal images are preprocessed with a haze-removal technique (2) a U-net segmentation network is utilized to classify pixels into background, artery or vein (3) a tracking algorithm is applied for vessel-wise classifications. The proposed method is tested on a clinical dataset and the results present an accuracy of 93.57% for vessel-wise classifications.