Discrimination And Realization Of Bifurcation And Intersections Points In Fundus Vascular Images

All the systemic diseases that affect vascular networks,such as diabetes,hypertension and arteriosclerosis are more and more general due to growing global wealth and an aging population.Therefore,the examination of blood vessels is critical to the diagnosis and prognosis of these diseases.The geometric properties of the fundus vasculature at the bifurcation and intersection,such as angle of intersection,vessel width,and bifurcation asymmetry,are indicators of whether the fundus vessels are healthy.At the same time,the blood vessel level corresponding to the position of the fundus abnormality can be used to provide an auxiliary basis for judging the severity of the disease according to the position of the bifurcation point,in general,the lower the level,the more serious the disease is.Moreover,the discrimination of the bifurcation and intersection is the basis for image registration based on feature or non-rigidity,image stitching,vascular grading and biometric security applications.Both the bifurcation point and the intersection point are key identification points in these cases.A bifurcation point is a point at which a blood vessel creates a bifurcation in circulation,and an intersection point refers to a point generated by projection of two non-intersecting blood vessels on a fluoroscopic or color map.The diagnostic information of the forked intersections is generally used based on accurately locating and distinguish them.Based on the above descriptions,the main work and achievements of this thesis are as follows: The domestic and international discrimination technology with development status of bifurcation intersection point for the fundus vascular image has been firstly analyzed,and the key techniques for the classification of bifurcation intersection points have been summarized.The second is the extraction of the bifurcation intersection in the fundus vascular image.The nodes have been detected by using convolution operations after the processes such as image preprocessing,blood vessel segmentation,image filling,breakpoint connection and bone formation.An improved method of merging erroneous bifurcation point based on the wall width of maximum vessel has been proposed.And all the information of location has been obtained eventually.Experiments results show that the improved method improves accuracy by 2% over current best results;the third is the classification of the bifurcation intersections in the fundus vascular image.A new algorithm of classification based on transfer learning for fundus bifurcation points and intersections has been proposed in this thesis.The traditional method and auxiliary adaptation of the data set are expanded after obtaining the data set,and a migration has been constructed based on neural network Inception-v3.The accuracy and efficiency of classification for the intersections have been improved after learning the network.And the accuracy of the classification for complex intersections has been ultimately improved by 3% over the latest results.The fourth is the operation of repair for the error point has been proposed in order to perform the extraction and classification of the bifurcation intersection.And the information of blood vessel circulation is combined to avoid the misjudgment caused by only the local features of the node.The algorithms and improvements presented in this thesis have been tested on both clinical medical images and public data sets,and several comparison methods were introduced.The proposed method for extracting node has been improved on several evaluation indicators compared with the current best results.And the proposed algorithm of classification at complex intersections has also been improved.After that,the research will focus on how to directly performing node discrimination on the original image without relying on the blood vessel segmentation.