Trainable V4-like filters for detecting retinal vascular bifurcations

Abstract

A poster presented by the authors at the AVA/BMVA SPRING (AGM) MEETING 2011 Cardiff University. We propose a computational model for shape-selective neurons in area V4 of visual cortex. Such neurons are known to respond to moderately complex stimuli that are local combinations of single-orientation stimuli. We implemented this model in filters that detect features which are composed of several lines of different orientations. The proposed filters are trainable, in that the specific feature to which such a filter responds is used to determine the structure of the filter. To configure such a filter we select given channels of a bank of Gabor filters and combine them using an AND-gate-like operation. Their selection is determined by the automatic analysis of a user-specified feature. Consequently, the configured filter responds to the same and similar patterns. We demonstrate the effectiveness of such filters by applying them on retinal fundus images to automatically detect the vascular bifurcations. The detection of such bifurcations is important for finding signs of various cardiovascular diseases. With only 25 filters we achieved a correct detection rate of 98.52% at a precision rate of 95.19% on 40 binary fundus images, containing above 5000 bifurcations manually annotated by the authors. Other authors (Bhuiyan et al, 2007, IEEE Conf. on Signal-Image Technologies and Internet-Based Systems, 711-718) report a detection rate of 95.82% on a smaller dataset of five retinal images. The novel automatic configuration gives an edge to our approach over other models regarding generalization ability. In principle, all vascular bifurcations can be detected if a sufficient number of filters are configured and used.