Blood Vessel Segmentation in Retinal Images Using Convolutional Neural Networks

Abstract

Vasculature extraction is an essential step in automated detection of retinal diseases. In spite of many approaches proposed for blood vessel segmentation, many of them are faltered in segmentation of fine blood vessels as they need low-level spatial information or these approaches tend to generate false positives at the branch ends of blood vessels. Occasionally over-segmentation and under-segmentation are resulting in spurious results in measuring width of blood vessels in quantitative studies. The proposed blood vessel segmentation method gives very promising results as it achieve a dice coefficient of 0.84 on STARE database and 0.844 on DRIVE database.