Multi-mode Fundus Image Registration Method Research

In order to have a comprehensive understanding of the fundus situation, doctor needs to integrate the information of the fundus images obtained from different imaging devices. Due to the rotation of the eye and head of the patient, there will be a spatial position deviation between the corresponding physiological structures in the fundus images. Fused the features of fundus image directly will result in the dislocation of the corresponding physiological structure. Therefore, images need to be registered before feature fusion. According to the shape and brightness of optic disc as well as stable vascular structures in fundus image, a registration method based on vascular and optic disc is proposed.Firstly, according to the characteristics of the optic disc and blood vessels in color and fluorescence angiography fundus images, Hough transform and morphological methods are used to locate optic disc in color fundus image and fluorescein angiography fundus image. According to the shape features similar to vascular branches and high contrast features, the multi-scale Hessian matrix and morphological methods are used to segment blood vessel of color fundus image and fluorescein angiography fundus image. Then SURF features are extracted to form SURF descriptors in the segmented blood vessel images. Secondly, the initial filtering and matching of feature points are realized by using the characteristics of Hessian matrix trace and Euclidean distance, and the matching points are purified by random sample consensus algorithm. Finally, the least square method is used to estimate the spatial transformation parameters and then the registration image is obtained by interpolation.Fundus images with different severity were experimented. The experimental results show that our method can realize the automatic registration of images with different resolutions and rotation angles. The average matching rate can be reached 93.3%and the root mean square error can reach the sub pixel level, which proves that the proposed method has a good accuracy and strong robustness.