A Hardware-implementable System for Retinal Vessel Segmentation GIOVANNI COSTANTINI, DANIELE CASALI, MASSIMILIANO TODISCO Department of Electronic Engineering University of Rome “Tor Vergata” ITALY Abstract: - A retinal vessel segmentation method based on cellular neural networks (CNNs) is proposed. The CNN design is characterized by a virtual template expansion obtained through a multi-step operation. It is based on linear, space-invariant 3×3 templates, and can be realized using real-life devices with minor changes. The proposed design is capable to perform vessel segmentation within short computation time. It was tested on a publicly available database of color images of the retina, using receiver operating characteristic (ROC) curves. The simulation results show the good performance, comparable with the best existing methods. Key-Words: - Cellular neural networks, line detection, vessel segmentation, retinal imaging 1 Introduction Automatic segmentation of blood vessels in retinal fundus images plays an important role in the diagnosis of several pathologies, like hypertension, diabetes, cardiovascular disease [1]. Several morphological features of veins and arteries (diameter, length, branching angle, tortuosity…) have diagnostic relevance. Accurate vasculature segmentation is a difficult task for several reasons: the presence of noise, the low contrast between vessels and background, the variability of vessel width, brightness and shape. Moreover, due to the presence of lesions, exudates and other pathological effects, the image may have large abnormal regions. Several methods have been proposed in the literature to address these problems, including matched filtering [2], tracking methods [3], multithreshold probing [4], and supervised classification [5]. One important aspect, rarely addressed, is the complexity of the algorithm and its efficient hardware implementation in order to reduce the time spent for the segmentation. One attractive paradigm for parallel real-time image processing is represented by cellular neural networks (CNNs) [6,7]. To the best of our knowledge retinal vessel extraction with CNNs has been previously proposed only in [8], where segmentation is obtained through histogram modification, local adaptive thresholding and morphological opening. This method presents two drawbacks. First, it relies on several design parameters: the scaling factors of local mean and variance (α and β), the neighborhood size, the structuring element for opening. Since no guidelines are available for their setting, they must be empirically tuned. Moreover, nonlinear CNN templates are required for local estimation of the variance. In this paper we present a new CNN based approach for detecting vessel pixels in color fundus images, avoiding the above mentioned disadvantages. The proposed method is based on the detection of linear structures through an operator introduced by Dixon and Taylor [9]. This operator, previously used in mammographic image analysis [10], could be realized with simple CNN templates. However, our simulations showed that to obtain state-of-the-art performance in retinal vessel segmentation the template size must be 15×15. So, to simplify the physical realization of the CNN, we adopt a multi-step operation with virtual template expansion. This network could reduce the segmentation time of some orders of magnitude with respect to a sequential digital realization. The CNN operation has been simulated on a digital computer and the corresponding performance has been evaluated on a publicly available database through receiver operating characteristic (ROC) analysis. The ROC curve shows the effectiveness of our method. The rest of the paper is organized as follows. In the next Section we illustrate the proposed method to detect vessel pixels. In Section III we describe the CNN realization. In Section IV some simulation results are presented and compared to existing techniques for vessel segmentation. Some comments conclude the paper. 2 Line Detection Usually in RGB non-mydriatic images the green channel exhibits the best vessel/background contrast while the red and blue ones tend to be very noisy. Therefore, we work on the inverted green channel images without performing any further preprocessing. An example is shown in Fig. 1: vessels appear brighter than the background. Grey level -1 corresponds to black, level +1 corresponds to white. The proposed method is illustrated in Fig. 2. The average grey level is evaluated along lines of fixed length l passing through the target pixel (i,j) at different orientations. We consider 12 orientations (15° of angular resolution). The line with the largest average grey level is found: its average value is denoted with Lij. The difference S ij = L ij – N ij represents the line strength of the pixel [10], where N ij is the average grey level in the square window, centered on the pixel, with edge length equal to l. LATEST TRENDS on COMPUTERS (Volume II) ISSN: 1792-4251 568 ISBN: 978-960-474-213-4