Multi-scale morphological analysis for retinal vessel detection in wide-field fluorescein angiography

We describe a new method for accurate retinal vessel detection in wide-field fluorescein angiography (FA), which is a challenging problem because of the variations in vasculature between different orientations and large and small vessels, and the changes in the vasculature appearance as the injection of the dye perfuses the retina. Decomposing the original FA image into multiple resolutions, the vessels at each scale are segmented independently by first correcting for inhomogeneous illumination, then applying morphological operations to extract rectilinear structure and finally applying adaptive binarization. Specifically, a modified top-hat filter is applied using linear structuring elements with 9 directions. The maximum value of the resulting response images at each pixel location is then used for adaptive binarization. Final vessel segments are identified by fusing vessel segments at each scale. Quantitative results on VAMPIRE dataset, which includes high resolution wide-field FA images and hand-labeled ground truth vessel segments, demonstrate that the proposed method provides a significant improvement on vessel detection (approximately 10% higher recall, with same precision) than the method originally published with VAMPIRE dataset.