Brain magnetic resonance image segmentation method based on scattering graph neural network

Abstract

The invention discloses a brain magnetic resonance image segmentation method based on a scattering graph neural network, and the method comprises the steps: employing super voxels as basic units, and generating the same number of super voxels for a reference image and a to-be-segmented image, so as to reduce the calculation complexity of a model; meanwhile, considering self information, surrounding neighbor information and spatial position information of the super voxels, and pre-extracting gray features, tensor features and key point spatial prior features of the super voxels; thirdly, due to the fact that certain topological structure information is implied among the super voxels of the brain, the super voxels serve as nodes to construct a topological graph, a scattering graph neural network is adopted to learn global topological information, and node features are updated; finally, feature matching is directly conducted on the super voxels of the image to be segmented and the super voxels of the marked reference image, a semantic segmentation result is obtained, the method can be well applied to the brain magnetic resonance image, and the tissue structure of the brain magnetic resonance image is effectively segmented.

Description

technical field [0001] The invention relates to a brain magnetic resonance image segmentation method based on a scattergram neural network, and belongs to the field of digital images. Background technique [0002] There are unstructured data in real life, such as social network, molecular structure and other graph structure data. For this kind of graph-shaped data, each node has a different topology (the number of neighbor nodes, the distribution of neighbor nodes, etc. are different), so unlike the convolutional neural network CNN, the graph convolutional neural network GCN does not With translation invariance, the convolution kernel with parameter sharing cannot be used. GCN obtains a new node representation by aggregating the features of each node and its neighbor nodes, and embeds the topology information between nodes when each node performs message transmission. However, GCN will strengthen the similarity between adjacent nodes when integrating graph structure inform...

AI Technical Summary

Problems solved by technology

[0004] At present, traditional image matching algorithms, such as grayscale-based template matching algorithms and feature-based matching algorithms such as SIFT and KLT, have been applied in many fields of computer vision and have achieved quite good results. The time complexity of the matching is very high, and it is not suitable for 3D magnetic resonance images with relatively large volume

[0005] To sum up, the deep learning model requires a large amount of pixel-level annotation data to obtain ideal results, and the segmentation and annotation cost of obtaining magnetic resonance images is very high, so image matching methods can be considered to reduce the annotation cost, with supervoxel as the basic unit, to reduce the matching complexity. At the same time, the current research has not yet disclosed how to use the topological structure information between supervoxels in the magnetic resonance image, and consider using the scatter map neural network to learn the topological structure between supervoxels, so as to be more comprehensive. supervoxel features

Images