Automatic detection of vascular bifurcations and crossings in retinal images using orientation scores

Several ocular and systemic diseases such as hypertension and arteriosclerosis cause geometrical and functional changes to the vasculature in retinal images, including alterations in the shape of vascular bifurcations and crossings. To use the diagnostic information of the junctions, it is important to detect them first. In this work, a novel BIfurcation and CRossing detection method using Orientations Scores (BICROS) is introduced. The Brain-inspired orientation score transformation lifts the image to the joint space of positions and orientations using directional anisotropic wavelets. Candidate junctions are selected based on their geometrical properties in this space. Then false detections are eliminated in a supervised manner. Additionally, a more conventional pipeline for junction detection based on morphological analysis of vessel segmentations is included. Finally, both approaches are combined and the resulting junctions are classified into bifurcations and crossings. The proposed method outperforms state of the art on a public and private dataset.