Blood Vessel Extraction in Low Contrast Fundus Images having Lesions under Diabetic conditions

Abstract

The significance of vessel segmentation approach is increasing in recent years. However, it poses various complex challenges which are difficult to handle and causes degradation in the efficiency of the existing models. Since Blood vessels share the same color and intensity information as that of dark lesions, it is mandatory to extract and remove blood vessels in automatic diabetic retinopathy detection. Thus, a high quality vessel segmentation technique is required to handle these challenges. This paper puts forward a segmentation algorithm to evaluate vessel and non-vessel regions of fundus images using high boost filtering and principal curvature analysis followed by morphological operation. Contrast limited adaptive histogram equalization and second order Gaussian filtering is introduced as an intermediate step that will improve over all detection efficiency. Accuracy of fundus images having inadequate contrast is enriched by color normalization followed by contrast enhancement. The proposed technique is evaluated using the DRIVE and MESSIDOR databases. Results from suggested methods are compared with results from prevailing methods to ascertain superiority in segmentation quality.