Comparison of Pixel and Subpixel Retinal Vessel Tree Segmentation Using a Deformable Contour Model L. Espona 1 , M.J. Carreira 1 , M.G. Penedo 2 , and M. Ortega 2 1 Computer Vision Group. Electronics and Computer Science Dpt. University of Santiago de Compostela. Spain luciaep@usc.es, mjose@dec.usc.es 2 VARPA Group. Computing Dpt. University of A Coru˜ na. Spain mgpenedo,mortega@udc.es Abstract. This paper presents a comparison of pixel and subpixel performance of the snake-based system designed to detect the vessel tree in eye fundus images. The automatic analysis of the retinal vessel tree facilitates the computation of the arteriovenous index, which is essential for the diagnosis and evolution of several eye diseases. A high accuracy is required to correctly assess the clinicians and it is insufficient when working at a pixel level. The developed model is inspired in the classical snake but incorporating domain specific knowledge and profits from the automatic localization of the optic disc and from the extraction of vascular tree centerlines previously developed in our research group [1]. The efficiency and accuracy of the detection of arteriovenous structures are evaluated using the publicly available DRIVE database and an equivalent system configuration for pixel and subpixel results. Results demonstrate that, although more time consum- ing, subpixel retinal vessel extraction is much more reliable, keeping relatively low values of computing time. Keywords: snakes, segmentation, retinal vessel tree, eye fundus image, subpixel. 1 Introduction The automatic analysis of blood vessels is becoming more and more important in many clinical investigations and scientific researches related to vascular features. The early diagnosis of several pathologies, such as arterial hypertension, arteriosclerosis or dia- betic retinophaty could be achieved analysing the vascular structures. The Digital Colour Fundus Photographs here used are a non invasive and innocuous technique to obtain the retinal vascular tree. Moreover, a specific CAD system is also necessary in large-scale ocular screening programs to make the ophthalmologist diagnosis process more efficient and accurate [2]. The retina arteriovenous index (AV index)indicates the relation between afferent and efferent blood vessels, that is arteries and veins of the retina. This index takes a vital priority in order to diagnose these illnesses and evaluate their consequences [3]. This paper deals with the selection of an appropiate precision level for our vascular tree detection system, evaluating its performance at pixel and subpixel level. This seg- mentation of the vascular tree would constitute the first step to allow the precise and robust AV index measuring. The eye fundus images are quite problematic 2-D medical images. The main dif- ficulties in them are the inadequate contrast, lighting variations and remarkable noise J. Ruiz-Shulcloper and W.G. Kropatsch (Eds.): CIARP 2008, LNCS 5197, pp. 683–690, 2008. c  Springer-Verlag Berlin Heidelberg 2008