Blood Vessel Extraction Adapted For Retinal Fundus Images With Pathologies

The retinal fundus image is an important basis for diagnosing of eyedisease.Automatic retinal image processing and analyzing system based on computerhas an important application value in auxiliary medical diagnosis. Retinal vasculatureis an important anatomical structure in retinal fundus image, which spread over wholeretinal image as well as tree network structure. Vessel extracting plays an importantrole in diagnosis of disease such as glaucoma, cataract, diabetics, hypertension, andarteriosclerosis, which often cause micro-vessel change. On the one side, manyfundus image analysis and identification algorithms need the information of vascularstructure in advance. On the other side, as a relatively stable physiologicalcharacteristic structure，vessel has important role in diagnosis system. For example, itsdistribution characteristics and feature information are favorable for the assessment ofretinal disease, it benefit for the localization of other anatomical structure such asoptic disc, fovea, and also for lesion detection and multi-source or multi-modal retinalimage registration and integration.A lot of research work of vessel extraction has proceeded in all over the worldand has made great progress. Although vessel extraction result is good in normalretinal images, it is difficult to extract vascular network structure in disease imagesaccurately, due to the edge of the optic disc, exudates or other lesions has a similarcharacteristics as vessel and the interference of dark gap between nearly multiplybright areas. This paper focus an in-depth study of vessel extraction in retinopathyimage.(1) Due to vascular-like characteristics existed in disease images and theinterference of retinopathy border, this paper proposes a vessel detection methodusing non-vessel boundary suppression based on Gabor filtering processing. Themethod is achieved by considering the maximum response properties of thesymmetrical Gabor filter function to the bar symmetric mode and asymmetric Gaborfilter function to the edge. In combination with non-maximum thinning andmulti-threshold technical based on standard hysteresis threshold binaryzation, theproposed method can keep complete vessels skeleton structure after rulings outremaining non-vessels pixels.(2) After Gabor filtering and multi-threshold processing, we also design aregional growth algorithm to realize the extraction of vessel through judging vascularwidth in vertical direction of seed point. The algorithm uses the response of Gaborfilter, while using vascular skeleton structure as seed of region growing. Based onregional growth, we then get more complete and smooth vessel structure bymathematical morphology, which can make up omission areas caused in the growthprocess.The proposed vessel extraction method mainly solves the interference of retinallesions and non-vessel structure in vessel extraction. Compared with similar methods,our vessel extraction method can obtain complete vessel structure accurately in disease retinal images.