Article Preview
Top1. Introduction
The automated segmentation of retinal vessels is the crucial stage in the diagnosis of many diseases such as hypertension (Leung et al., 2004), obesity (Mitchell et al., 2005), glaucoma (Wang et al., 2006) and diabetic retinopathy (Morello, 2007). These diseases often result in changes on bifurcations and tortuosity of retinal vascular. Hence, analyzing vessel features gives new insights to diagnose the corresponding disease early (Vijayakumari & Suriyanarayanan, 2012). Also for diagnosing the disease progress of certain patient along time, the blood vessels segmentation is necessary in automated registration of two retinal images of that patient at different times (Khan et al., 2011). Since the manual segmentation requires training and skill extensive research efforts have been devoted to automating the segmentation (Fraz et al., 2012a). Nevertheless, reliable vessel segmentation faces several challenges (You et al., 2011): “1) Retinal vascular has variant widths, lengths, bifurcations and tortuosity; 2) the narrow vessels usually may be lost by segmentation since they disappear among various local surroundings; and 3) various structures appear in the retinal image disrupt segmentation such as optic disc, fovea and exudates”.
This article proposes two improvements of the previous baseline approach (Asad et al., 2012a) used for automatic segmentation of blood vessels in retinal images based on the ant colony system (ACS) (Dorigo & Gambardella, 1997). The first improvement is in features where the length of previous features vector used in segmentation is reduced from eight to five since four less significant features are replaced by a new more significant feature when applying the correlation-based feature selection heuristic (CFS) (Hall, 2000). The second improvement is in ACS where a new probability-based heuristic function is applied instead of the previous Euclidean distance based heuristic function. Since the large number of computed features increases the classification complexity, time and reduces accuracy. So that, feature selection is an essential step for successful classification because it removes irrelevant features and achieves less complex, more accurate and faster classification. In this paper, CFS is used and it recommended the best feature vector consisting of five features out of fifteen features for segmentation. The new performance of this improved approach is evaluated on a publicly available database of retinal images for scientific research DRIVE (Staal et al., 2004) in terms of the sensitivity, specificity and accuracy.
The rest of this article is organized as follows: Section 2 surveys the previous popular related work. Section 3 gives scientific background of the used features, CFS and ACS. Section 4 presents the previous baseline approach and its two improvements. Section 5 reports the results of experimental evaluation of improved approach. The conclusions and future work are finally presented in Section 6.