Automatic selection of measurement points in multispectral fundus images for retinal oximetry

Abstract

Abstract: : Purpose: To develop a method for automatic selection of measurement points using morphological filters for automatic evaluation of oxygen saturation in the retina using a reflectance fundus oximeter. Methods: We constructed morphological filters designed to enhance vessel like structures. A skeleton of a thresholded fundus image was then used to locate the arterioles and venules. Measurement points were then selected inside the vessels at the darkest spot to reduce the effects of specular reflection and corresponding points are selected on the retina just outside the blood vessel. The accuracy of the method was evaluated in 15 fundus images from 5 eyes and compared with a human observer. Results: The method is able to select blood vessels in fundus images with high accuracy. 96±4 % (mean±standard deviation) of blood vessels 100–150 µm in width were detected, 82±8% of blood vessels 70–99 µm in diameter were detected and 43±14% of blood vessels 30–69µm. False positive findings are easily detected and removed. Conclusions: Although still not as accurate as a human operator the software has the potential to automate and greatly speed up the processing of retinal images.  

Keywords: imaging/image analysis: clinical • blood supply • diabetes