We employed multiscale line operators (MSLO) in order to segment blood vessels in digital fundus images. Separately, a median filter technique was used in order to provide results that were compared to those of the MSLO. The green channel of the colour image was used, and both sets of results were further enhanced by subsequently employing a simple ‘‘randomly seeded’’ region-growing algorithm. We applied this approach to two sets of retinal images, namely, the ARIA (www.eyecharity.com/aria_online/) and STARE (www.ces.clemson.edu/ahoover/stare/) retinal image archives. The ARIA dataset contained colour fundus images from healthy subjects, diabetic subjects, and age-related macular degeneration (AMD) subjects. Similarly, the STARE dataset contained images from both ‘‘normal’’ (i.e., healthy) and ‘‘abnormal’’ (i.e., diseased) eyes. Manual segmentations of the blood-vessel structure for all images in the ARIA and STARE datasets were obtained by a retinal image interpretation expert. These images were then taken to be our gold standards. Receiver-operator characteristic (ROC) curves were determined and the areas under the ROC curve (AZ) were obtained. A large increase in efficiency for our MSLO algorithm was observed for the entire datasets (ARIA AZ=0.899; STARE AZ=0.953) compared to basic thresholding alone (ARIA AZ=0.608; STARE AZ=0.753). Interestingly, the simple median filter algorithm followed by region growing also performed well (ARIA AZ ¼ 0.888; STARE AZ ¼ 0.947). Our results compared favourably to those results of previous segmentation procedures for the STARE dataset. As expected, the best results were found for the healthy control group for ARIA and for the normal subjects for STARE.
- retinal image processing
- automated blood vessel tracing
- diabetic retinopathy
- age-related macular denegeration