Unsupervised Change Detection in Multitemporal Images of the Human Retina

Abstract

Diabetes is a growing epidemic in the world, due to population growth, aging, urbanization, and increasing prevalence of obesity and physical inactivity. Diabetic retinopathy is the leading cause of blindness in the western working age population. Early detection can enable timely treatment minimizing further deterioration. Clinical signs observable by digital fundus imagery, include microaneurysms, hemorrhages, and exudates, among others. In this chapter, a new method to help the diagnosis of retinopathy and to be used in automated systems for diabetic retinopathy screening is presented. In particular, the automatic detection of temporal changes in retinal images is addressed. The images are acquired from the same patient during different medical visits by a color fundus camera. The presented method is based on the preliminary automatic registration of multitemporal images, and the detection of the temporal changes in the retina, by comparing the registered images. An automatic registration approach, based on the extraction of the vascular structures in the images to be registered and the optimization of their match, is proposed. Then, in order to achieve the detection of temporal changes, an unsupervised approach, based on a minimum-error thresholding technique, is proposed. The algorithm is tested on color fundus images with small and large changes.