Retinal vasculature segmentation by morphological curvature, reconstruction and adapted hysteresis thresholding

Automatic retinal blood vessel extraction is very important for early diagnosis and prevention of several retinal diseases. In this paper, a new retinal vasculature segmentation algorithm is proposed based on mathematical morphology, principal curvature, non-maximal suppression and hysteresis thresholding based morphological reconstruction. The blood vessels are enhanced by applying the top-hat transformation and computation of maximum principal curvature at multiple scales. Vessel centerlines are then obtained by non-maximal suppression followed by adapted hysteresis thresholding and morphological reconstruction. The principal curvature image is double thresholded and morphologically reconstructed to generate the vessel skeleton map which is the aggregate threshold for region growing of detected vessel centerlines to obtain the segmented retinal vasculature. The proposed method is evaluated using the images of two publicly available databases, the DRIVE database and the STARE database. Achieved average accuracy for DRIVE and STARE is 0.9419 and 0.9434 respectively. Experimental results show that the proposed algorithm is comparable with other approaches in accuracy, sensitivity and specificity.