Research On Vessel Extraction Method Of Diabetic Retinopathy Fundus Image

The retinal vessels are the deep microcirculation network that are the only direct observation without wound, and the diameter, the bending degree of vessel and the bending degree of bifurcation angle are the important diagnosis index of diabetes, hypertension, cardiovascular and cerebrovascular and ophthalmology diseases, so the effective segmentation method of retinal vessels can assist the doctor to early diagnosis and treatment of the above cases. Due to the restrictions of many non-ideal imaging conditions and physiological structure, direct vessels segmentation is difficult. Therefore, the retina image must be effectively enhanced to remove the impact of disc bright before vessel segmentation. The lesion parts need to be detected and eliminated to obtain accurate segmentation results when vessel is segmented.A novel enhancement method of retinal image based on the multi-scale and multi-direction is proposed, according to the morphology structure of the vascular. Firstly, the green channel of retinal image is obtained by image preprocessing. Then the green channel image is processed respectively with multi-scale filter based on Hessian matrix and multi-directional matched filter. Finally, the maximum value in each pixel between multi-scale and multi-dimensional filter images is chosen as the ultimate enhanced results. A vessel segmentation method based on global and local thresholds is proposed based on the results of image enhancement. The main vessels and small vessels are respectively segmented through the global threshold of two-dimensional maximum entropy and the local threshold of moving average. Finally, the vessel segmentation results are obtained via analyzing the region connectivity and combining the main vessels with small vessels. Only using the segmentation method proposed in this paper cannot extract the precise vascular network because of the bright of optic disc and the lesion region exist in the retinopathy images. Therefore, a novel vessel extraction method of diabetic retinopathy based on the optic disc localization and lesion detection is proposed. Firstly, the accurate disease detection result is obtained by a series of morphological operations, and then removed from the vessels segmentation results. Finally, the final network of blood vessels is obtained via connecting the broken vessels in the area of lesions according to the principle of gradient intensity and direction similar.In this paper, vessel segmentation methods are tested in the public database of Drive and the public database of Stare, and the results are compared to the expert manual segmentation results. A large number of experimental results show that the method can achieve higher accuracy, get a better segmentation results.