Research On MR Image Segmentation Based On Supervoxel Matching

Magnetic resonance imaging(MRI)is a method that uses radio frequency pulse to excite the hydrogen nucleus in the magnetic field,and then uses the induction coil to collect the signal for image reconstruction.Magnetic resonance imaging has been widely used in clinical examination and disease diagnosis because of its small radiation and good resolution of soft tissue.Accurate automatic segmentation of MRI images is of great significance for clinical diagnosis and scientific research.In recent years,deep learning technology has made remarkable achievements in the field of image processing,and has also been widely studied and applied in medical image processing and analysis.In many medical image processing tasks,the deep learning method is superior to the traditional algorithm,and the fully convolutional neural network has become the mainstream method of image semantic segmentation.Although the effect of this kind of method is very good,they need a lot of labeled data as supervision information to guide neural network training.In the case of insufficient labeled data,deep learning method is difficult to obtain good results.Medical image annotation needs experts with professional knowledge to complete,so the cost of label acquisition is very high.Considering the high similarity of the structure of the same tissue between individuals,this paper proposes a segmentation method of magnetic resonance image based on supervoxel matching.Only one data with segmentation label is needed as the reference image,and semantic segmentation results of other images is obtained by matching.The framework includes four modules: supervoxel generation,extract supervoxel’s feature,update supervoxel’s feature and supervoxel feature matching.Firstly,a consistent number of supervoxels are generated for the reference image and the image to be segmented.Supervoxels are image regions composed of adjacent pixels with similar features.Compared with pixel by pixel matching,the input scale of the algorithm can be reduced based on supervoxel matching,and then reduce the processing time of the algorithm.Secondly,the features of supervoxels are extracted.In this paper,the gray histogram features,tensor features and spatial prior features of key points are extracted respectively.The histogram describe the intensity distribution of the supervoxel itself,tensor describe the neighborhood information of the supervoxel and spatial prior features of key points describe the global position information of the supervoxel.Thirdly,the supervoxels of the reference image and the image to be segmented are used as nodes to construct a graph,and the scattering graph convolutional neural network is used to learn the supervoxel’s feature.Finally,the learned features are used for supervoxel matching,the category of supervoxels of the image to be segmented can be obtained,the final image semantic segmentation result can be obtained by mapping supervoxel-level category labels to voxel-level.In order to evaluate the performance of the proposed MRI segmentation method,this paper conducted experiments on two public datasets IBSR18 and MRBrain S18.The proposed method is verified by qualitative comparison and quantitative analysis,and compared with other segmentation methods.Experiment results show that the proposed method can achieve good segmentation results in the case of only one labeled image.