The 3 rd European Medical and Biological Engineering Conference November 20 – 25, 2005 EMBEC'05 Prague, Czech Republic IFMBE Proc. 2005 11(1) ISSN: 1727-1983 © 2005 IFMBE A RIBBON OF TWINS FOR EXTRACTING VESSEL BOUNDARIES B. Al-Diri and A. Hunter The University of Lincoln baldiri@lincoln.ac.uk Abstract: This paper presents an efficient model for automatic detection and extraction of blood vessels in ocular fundus images. The model is formed using a combination of the concept of ribbon snakes and twin snakes. On each edge, the twin concept is introduced by using two snakes, one inside and one outside the boundary. The ribbon concept integrates the pair of twins on the two vessel edges into a single ribbon. The twins maintain the consistency of the vessel width, particularly on very blurred, thin and noisy vessels. The model exhibits excellent performance in extract- ing the boundaries of vessels, with improved robust- ness compared to alternative models in the presence of occlusion, poor contrast or noise. Results are pre- sented which demonstrate the performance of the dis- cussed edge extraction method, and show a signifi- cant improvement compared to classical snake formu- lations. Introduction Vessel segmentation algorithms are a critical compo- nent of blood vessel analysis systems. The segmentation of vascular structures plays an important role in diagno- sis, surgery and research, in many important systematic diseases such as diabetes and hypertension. Furthermore, the vascular tree seems to be the most appropriate repre- sentation for image registration applications. A number of authors have investigated the use of active contour models in retinal vascular segmentation. However, none have reached the level of performance re- quired for robust automatic detection of the entire net- work of blood vessels, which is needed for measurement techniques used for identifying and grading the sever- ity of diabetic retinopathy. We have thus introduced a new parametric active contour model called a ribbon of twins. our model uses the Gradient Vector Flow method – which provides good performance through concavi- ties and noise features; the “twin” method – which uses two contours coupled by spring models to overcome ini- tialization and localized feature problems; and the “rib- bon” method, which couples two snakes with a consis- tent width parameter. Our model integrates these differ- ent features to produce a novel hybrid model. Parametric contours In the parametric approach, an active contour [1] is represented as a curve or spline, v(s)=(x(s), y(s), where v is a “vertex” and x and y represent the coordinates of the vertices and are functions of the normalised arc length 0 s 1. The active contour has a dynamic behaviour that deforms from an initial position and hopefully con- verges to the boundary of the object. The energy func- tional E snake composed of energy terms defines this be- haviour: E snake = 1 0 E snake (v(s))ds (1) = 1 0 E int (v(s)) + E pho (v(s)) + E ext (v(s)) ds (2) – where E int is the internal energy term that is based on the curve itself, which is designed to keep the model smooth during deformation; E pho is the photometric en- ergy term that arise from the image information, which is defined to move the model toward an object boundary or other desired feature within an image; E ext is the external energy term that is proposed to improve the capture range of the photometric force. Internal energy The internal energy consists of a first order term and a second order term [1]. E int = α (s)|v (s)| 2 + β (s)|v ′′ (s)| 2 2 (3) where v (s) and v ′′ (s) denote the first and second deriva- tive respectively, and parameters α and β are the coeffi- cients of the internal energy term and represent tension and rigidity, respectively. Photometric energy The photometric energy is derived from the image to attract the snake toward desired objects such as bound- aries. Given a grey-level image I (x, y),the photometric energy leading an active contour towards edges is defined by [1]: E (1) pho = −|I (x, y)| 2 (4) E (2) pho = −|(G σ (x, y) I (x, y))| 2 (5) where G σ (x, y) is a two-dimensional gaussian filter with standard deviation σ , is the gradient operator, and is the 2D image convolution operator. The blurred gradi- ent based approach has a number of limitations. First,