Abstract: The distribution of feature points extracted by traditional fundus image registration methods is too dense, which leads to the inaccurate alignment of the images to be registered images. And the retinal vessel bifurcation feature points have sparse distribution and stable features to improve the accuracy and speed of image registration. Therefore, this paper proposed a deep learning-based fundus image registration framework for vessel segmentation and bifurcation feature point extraction. This framework is composed of two deep convolutional neural networks. The first is the fundus blood vessel segmentation network SR-UNet, which combines channel attention (SE) and residual blocks on the basis of U-Net to segment retinal vessels to assist feature points extraction. The second is the feature point detection network FD-Net, which is used to extract bifurcation feature points from the vessel segmentation map. The proposed registration model is tested on the public fundus registration data set FIRE. The correct matching rate of the feature points is 90.03%. Compared with more advanced retinal image registration algorithms, the algorithm proposed has better performance and strong robustness both in registration quantitative and visual analysis.

Key words: retinal image registration, deep learning, vessel segmentation network, feature point detection network, U-Net