Comparison of algorithms for segmentation of blood vessels in fundus images

We compare the performance of three segmentation algorithms, initially developed by the team to extract the blood vessel network in fundus images of infant retina, on publicly available data sets. We envisage to parametrize any data-dependent factor (such as thresholds) to adapt to the data set. Experimental results show that through the proposed supervised-learning approach to tune parameters, all three segmentation methods can be adapted to work satisfactorily on fundus images of adult eyes that are clinically healthy as well as those afflicted with various pathologies. In particular, a supervised learning method of parameter-tuning for a segmentation method based on the matched filter response technique is seen to perform consistently well across the three data sets to give an average sensitivity of 73.16% and specificity of 90.57%.