Research Of Medical Image Registration Based On Semantic Feature And Intensity Adaptation

Image registration is a key technology in the field of computer vision.Meanwhile,medical image registration is an essential part of computer-assisted image processing.With the development of deep learning,accurate registration results can be obtained by extracting image features and performing feature fusion through the convolutional neural networks.However,there are some problems with the existing registration methods,such as poor registration accuracy,inappropriate selection of similarity metrics,unrealistic modal translation,and the inability to general registration.In this paper,we explore the contribution of semantic feature to structural registration.In addition,we reduce the difference in appearance between images through an intensity adaptation and achieve the general parallel registration by combining the semantic feature and intensity adaptation.The main research contents of this paper are as follows:(1)In order to select an appropriate similarity metric and improve the registration accuracy,this paper explores the facilitation of feature alignment on intensity image alignment,presenting the semantic feature-based multi-stage network(SFM-Net).Specifically,we design a two-stage training strategy,the intensity image registration stage and the semantic feature registration stage.The former is for valid semantic feature learning and intensity-based coarse registration,while the latter is for semantic feature similarity metric constructing,achieving fine transformation of the anatomical structure.The same structure of both stages is composed of a dual-stream feature extraction module(DFEM)and a refined deformation field generation module(RDGM).DFEM constructed by dual-stream U-Net structure can capture semantic information in the decoder feature for structural alignment.Meanwhile,RDGM can generate accurate multi-scale deformation fields by performing a coarse-to-fine registration within a single network.Experiments on 3D brain MRI and liver CT datasets confirm that our proposed model achieves accurate and diffeomorphic registration results,outperforming other state-of-the-art methods.(2)In order to improve the registration of multi-modal medical images,this paper decouples the difference between images into the appearance and structural differences,presenting the intensity adaptive modal translation-based multi-stage network(IMM-Net).This hierarchical model comprises three networks: a coarse and a fine registration network,and a modal translation network.First,the coarse deformation field is learned through the coarse registration network,which is then utilized as structure-preserving information in the next modal translation network.The modal translation result is obtained through the additive intensity adaptation in the modal translation network.Then,this translated image is fed into the fine registration network as enhancing information to the original images,deriving a fine deformation field.We compose the coarse deformation field and the fine deformation field to deliver the final deformation field,achieving the mutual promotion of modal translation and multi-modal registration.Experiments on two 2D multi-modal brain image datasets confirm that our proposed model makes great progress in both modal translation and multi-modal registration compared with the existing methods.(3)In order to achieve a general registration,this paper focuses on the advantages of the two networks mentioned above,proposing the semantic feature and intensity adaptation-combined general parallel network(SIGP-Net).We apply the two-stage training strategy of the intensity image registration stage and the semantic feature registration stage.In the same structure of both stages,we change the classical serial strategy of affine transformation and deformable registration,designing a multi-task network with parallel learning of affine matrix,deformation field,and intensity adaptation mask to reduce the registration time.What’s more,we propose an attention-gating module and an intensity gradient loss function to maintain the registration accuracy.Experiments on 3D brain MRI,3D liver CT datasets and two 2D multi-modal brain image datasets verify that our proposed model achieves registration with high accuracy in both uni-and multi-modal images.