Purpose
Segmentation and localization of retinal structures is an essential pre-processing step for many methods in automatic or computer aided medical diagnosis. In this work, we present a framework for segmenting and localizing three important retinal structures in color fundus images: vascular tree, optic nerve head, and fovea region.
Methods
First we perform a vessel segmentation to extract the vascular tree. In this method a resolution hierarchy and a vesselness feature extractor is combined with hysteresis thresholding to generate a binary vessel map. As second step the Fast Radial Symmetry Transform is applied on the fundus image. This method is designed to localize the center point of circular structures, which we use to find the optic nerve head center and estimate its diameter. The obtained vessel mask and optic nerve head information is then used to estimate the center of the fovea region in the image by fitting a double parabola onto a calculated vessel density map through the optic nerve head. This estimation is then refined by analyzing a region of interest around it. Each method in the framework is tested on the public available high resolution fundus (HRF) database (see: http://www5.cs.fau.de/research/data/fundus-images/), and the results are compared to a gold standard. This database contains 15 images each of healthy eyes, glaucomatous eyes and eyes with diabetic retinopathy. The gold standard of an image contains the following information to evaluate the automatic results: -manually labeled binary vessel images -coordinates of the center and the diameter of optic nerve head -coordinates of fovea center
Results
The vessel segmentation shows accuracy over 0.93, the optic nerve head localization has an accuracy of 0.98. The macula localization is tested on a subset of 20 images containing both healthy and pathologic images. Until the presentation of this work we will continue with the evaluation using the other 25 images. This preliminary evaluation shows a localization error under 0.15 optic nerve head diameter.
Conclusions
Each of the presented methods in this framework shows a high accuracy in our evaluation. Thus, the framework can be used effectively to aid medical diagnosis by providing segmentation and localization of important retinal structures in fundus images.
Keywords: 551 imaging/image analysis: non-clinical