A Study Of The Optic Disc Detection Method Combined With The Improved GBVS Model And Fundus Vascular Structure Features

Color fundus image processing and analysis can not only be used to diagnose certain eye disease,also can help the doctor to diagnose some systemic diseases such as diabetes,high blood pressure,and track the development of the disease.An optic disc(OD)is the origin of the retinal blood vessels,it generally appears as an approximately circular and bright yellowish spot in the retinal fundus image,there are a large amount of thick blood vessels converged in its internal.Accurate OD localization can be used to locate the important anatomical structures such as the blood vessels and macula,and also can help to locate other lesions such as exudates and micro-aneurysms.Moreover,many characteristics(such as shape and size)of OD are auxiliary parameters in clinical diagnosis.Overall,OD localization is of great significance for fundus image analysis.Making full use of vessel structure and three features of intensity,contrast and phase congruency,an OD detection method is presented,which combined modified graph based visual saliency(Gbvs)model and structure feature of retinal blood vessels.Firstly,the images are preprocessed.Then,the Gbvs model is improved by changing the three features of color,intensity and orientation into intensity,contrast and phase congruency(PC),which accords with human visual perception characteristics,and the modified Gbvs model is used to obtain saliency map.Then,vein vessels are extracted and the least square is applied to fit parabolic,accurate localization of OD is obtained through comparing the size of the average significance between the parabola vertex neighborhood and the whole fundus image.Finally,the main blood vessels around the OD vicinity are removed and the boundary of the OD is detected using C-V model.The proposed method is tested on four publicly available datasets(DRIVE,MESSIDOR,STARE and DIABETEDO),and the average test results are 100%.99.25%?93.8%?97.7%,which are higher than that of existing representative methods.