Research And Application On CNN-based Brain MR Registration Method

Image registration is an image processing technology that aligns one image to the space coordinates of another image,which play an essential role in medical image processing.With the development of deep learning and computer vision,the research on image registration has also made breakthroughs.However,there are some challenges in the inter-subject brain registration based on CNN.Firstly,there are certain differences in the structure of different human brains,and the sizes and shapes of tissues of the brain are different.Because the semantics of the medical images is simpler than the natural images,which make U-Net difficult to learn the mappings among different areas and different tissues.Additionally,it is important to balance the registration accuracy and the smoothness of the deformation field.To overcome the above problem,this paper proposes the following solutions,mainly including:(1)This paper proposes a multi-scale registration network named AMS-Reg Net,which is based on U-Net like network and enhances the learning of the weight the mapping between pixels in different regions by adding Attention-Gate to the skipconnection.The network perform multi-scale feature extraction on the deep feature map to improve the registration accuracy for tissues of different sizes and shapes.(2)Based on the idea of cascading and reversible consistent registration,combined with AMS-Reg Net,this paper propose a cascaded multi-scale reversible registration network CMSR-Net.This network uses a U-Net-like network as a coarse registration network for preliminary registration of image pairs.And cascade a reversible consistency registration network based on AMS-Reg Net,and perform reversible registration training on it,so that the network can improve the smoothness of the deformation field and reduce the generation of folding regions while maintaining high registration accuracy.(3)Based on the above-mentioned registration algorithm,a visual brain MR imageassisted registration system is constructed.The system integrates brain MR image registration,medical image processing and result display,simplifying the operation and process of doctor registration,and assisting doctors in quickly registering the brain MR image.