Medical Image Segmentation Method Based On Message Passing Network

Medical images can provide doctors with visual imaging support to aid in the diagnosis,treatment and subsequent medical observation of diseases.Medical image segmentation technology likewise plays a crucial role in the size measurement,positioning,and shape of lesions.In the past,manual segmentation of medical images was a very tedious and time-consuming task,and operators would carry their own subjective consciousness,which could result in segmentation errors.In recent years,automation segmentation of medical image has become a hot research topic.Segmentation methods based on convolutional neural networks are often limited to the size of convolutional kernels,which can only extract local features of medical images.However,some human tissues in medical images,such as cerebrospinal fluid,are distributed on pixel points at long distances in the image,and thus long-distance feature interactions between these regions on the image are needed to be explored.For the traditional convolutional neural network is difficult to achieve the long-range dependence in medical image segmentation,in this paper,two medical image segmentation methods are established with the main research line of constructing interpretable learning methods for long-range message passing in feature space.(1)Medical image segmentation method based on feature space message passing network.(2)Medical image segmentation method based on feature space coarse and fine information passing.For the model to learn long-range features in medical image effectively and to enhance the ability of the model to recover long-range feature recovery in the up-sampling phase,this paper proposes a medical image segmentation method based on feature space message passing network,which consists of two kinds of modules,namely,dynamic message passing module and graph skip connection module.Firstly,the dynamic message passing module dynamically constructs a graph structure with voxels as nodes and voxel feature distances as node connection metrics at different layers of the encoding path of the model,and in particular,using voxels directly as nodes in the constructed graph structure solves the problem of mapping images in Euclidean space to the topological graph in nonEuclidean space in most tasks of image and graph association.On the constructed graph,the message passing operation based on the spatial-domain graph convolution is used to aggregate the node’s own features and neighboring features to achieve long-range feature interaction on the image.Meanwhile,the local features extracted with the convolution are fused to obtain more comprehensive contextual information.Secondly,the graph skip connection module passes the graph structure and long-range features on the encoding path to the corresponding layer of the decoding path to assist the recovery of long-range information from the decoding path.The experimental results show that the method achieves excellent segmentation results on several medical image datasets and outperforms other methods in terms of three metrics,including Dice Similarity Coefficient(DSC),Absolute Volume Difference(AVD)and Intersection over Union(Io U).In order to compensate for the high-frequency information component in the graph,i.e.,the detail information of the image,which is neglected by the graph convolution smoothing operation during the long-range feature learning,a medical image segmentation method based on feature space coarse-fine information passing is proposed in this paper,which consists of three parts,namely,coarse information extraction,fine information extraction and coarse-fine information interaction.Among them,the coarse information is composed of local features extracted by convolution and long-range features extracted by graph convolution.In particular,the graph structure used to extract long-range features is constructed using a channel-based attention strategy and the graph structure is obtained by adaptive learning of the model,so it has a strong generalization capability;the fine information is obtained using a graph wavelet scattering operation to obtain high-frequency features on the graph structure,and the nodes and voxels on the graph correspond one-to-one,then they are also represented as detail features on the image;the interaction of coarse and fine information is an adaptive way to capture the potential relationship between coarse and fine information using attention mechanisms.The experimental results demonstrate that the method plays a positive role in extracting detailed features for medical image segmentation,with significant improvement in results compared with other related methods and a more reasonable number of parameters in the model.