Multi-scale segmentation of retinal images

Abstract

Tiny bright golden patches can be seen on the dark retinal background of some carriers of X-linked retinitis pigmentosa, an incurable blinding disease. We are interested in analyzing quantitatively these unique "tapetal-like reflex" patches in order to increase our understanding of the cellular mechanisms of the disease. In this paper, we describe the multi-scale thresholding method and its application to the segmentation of the tapetal-like reflex in high resolution digital fundus images. Multi-scale thresholding is a local thresholding method that generates results very similar to that of human intuition. Unlike other local thresholding methods, our method successfuIly ignores small artifacts in dark regions and simultaneously generats high resolution definitions of objects. Other segmentation applications where there are many bright objects on a darker background should profit from the use of multi-scale thresholding.