Deformation Medical Image Registration Algorithm Based On Deep Prior Optical Flow Network

Abstract

Deep convolutional networks have become a common tool for image generation and reconstruction. Unlike people who attribute their excellent performance to the prior information that they can learn from a large number of image samples, the author of Deep Image Prior (DIP) believes that the generator The network can capture a large number of low-level image statistical information before any learning, which means that this information may not be learned through a large number of data sets, and it verifies that the CNN network has a better capture of natural image distribution information. Ability to imitate. Inspired by the idea of DIP, can the deep convolutional network's ability to capture a large number of low-level image statistical information be used in image registration to capture the deformation field between the floating image and the target image to compensate for the traditional manual features and the common convolutional network features Limitations of information capture capabilities. Because the traditional optical flow registration method only relies on gray information and gradient-driven deformation registration, the registration result is not accurate enough. Therefore, we propose an optical flow medical image registration method based on the depth image prior. The generator network extracts feature information to generate the deformation field, and the discriminator of the GAN network is used to further accurately improve the deformation field to make it closer to the real deformation field. Experimental results show that our proposed algorithm can effectively improve the registration accuracy, and its registration accuracy is better than Demons algorithm, medical image professional registration software Elastix and Scale Invariant Feature Transform (SIFT) Flow algorithm.