LAUSR

Arteriovenous nicking (AVN) is a prominent biomarker in assessing cardiac risk factors such as hypertension, atherosclerosis and coronary artery diseases. The retinal junctions play a vital role in AVN quantification and assessment. The human retinal vasculature is a complex tree structure of blood vessels originating from the optic disc and spans over the retina. As the retinal vasculature is multidimensional in nature, identification of retinal junctions at multiple scales is a challenging task. In this paper, to address this multi-scale junction detection and classification, a novel deep learning approach has been proposed using a region-based convolutional neural network (RCNN). The proposed method detects the accurate location of junctions using the junction proposal network (JPN) and classifies the junctions as cross-overs and bifurcations using the junction classification network (JCN). Compared to peer’s work, the proposed method detects multi-scale and immediate junctions. The method also estimates the junction area, which is another quantitative measurement of AVN. The precision, recall scores are evaluated on two publicly available datasets, DRIVE and IOSTAR. The results obtained are analyzed and compared with peer’s work.