Study On Three - Dimensional Reconstruction Of Fundus Fundus Image

Three-dimensional visualization of vascular structures can provide straightforward information of the morphology of vessels, spatial relations among these vessels and other relevant anatomic structures. Thus it is of vital importance in diagnosis and treatment of the fundus vessels.In this thesis, the 3D retinal vessel model is reconstructed by the 2D fundus image. Firstly, a kind of new 3D reconstruction method of fundus based on the anatomical structure of eyeball and the eye model is proposed. Firstly, the posterior pole of eyeball is approximated as the partial sphere by analyzing the physical structure of eyeball and the eye model. The partial sphere is calculated by fundus camera parameters and the axial length of eyeball, and the 3D model of fundus is reconstructed from the fundus image. Secondly, the retinal vessel segmentation method is proposed in the base of the green channel of the color fundus image, image enhancement is by contrast limited adaptive histogram equalization, and the central reflex is smoothed by anisotropic diffusion. After that, vessel is matched by the multi-directional adaptive difference of gaussian filter, then binarized vessel is obtained by global threshold, and lessions are removed by post-processing. At last, vessel parameters as center line and radius are calculated based on vessel segmentation results, then the 3D point cloud data of retinal vessel is constructed combining 3D model of fundus and vessel parameters, and the 3D surface of the vessel is reconstructed using tessellation method.The method of retinal vessel 3D reconstruction is established through the study of this paper. The 3D vessel model can be reconstructed by the 2D fundus image. Experimental results show that, the approach in this paper can reduce the difficulty of image acquisition and the effects of central reflex of vessel to improve the recognition rate of retinal vessel 3D reconstruction. Experimental results show that the proposed vessel segmentation method can produce accuracy of 95.8% for DRIVE dataset and 95.69% for STARE dataset.