Method and system for brain tumor segmentation in multi-parameter 3D MR images via robust statistic information propagation
View Patent ↗A method for brain tumor segmentation in multi-parametric 3D MR images. The method comprises: pre-processing an input multi-parametric 3D MR image; classifying each voxel in the pre-processed multi-parametric 3D MR image, determining the probability that the voxel is part of a brain tumor, and obtaining an initial label information for the image segmentation based on the classification probability; constructing a graph based representation for the pre-processed image to be segmented; and generating the segmented brain tumor image using the initial label information and graph based representation. This method tries to exploit the local and global consistency of the image to be segmented for the tumor segmentation and can alleviate partially the performance degradation caused by the inter-subject image variability and insufficient statistical information from training.
1. A method for brain tumor segmentation in multi-parametric 3D magnetic resonance (MR) images, comprising:
pre-processing an input multi-parametric 3D MR image;
classifying each voxel in the pre-processed multi-parametric 3D MR image, determining and assigning a classification probability that the voxel is part of a brain tumor;
labeling voxels if their classification probabilities are higher than a selected threshold;
obtaining initial label information for image segmentation using only labeled voxels;
constructing a graph based representation for the pre-processed image to be segmented; and
generating the segmented brain tumor image using the initial label information and graph based representation.
2. The method of claim 1 , wherein
classifying each voxel in the image to be segmented and determining its corresponding classification probability using a support vector machine (SVM) based classifier.
3. The method of claim 1 , wherein
constructing a feature vector for each voxel based on the multi-parametric image intensity information of voxels in its spatial neighborhood.
4. The method of claim 3 , comprising classifying the feature vector of each voxel using the SVM classification method.
5. The method of claim 1 , wherein constructing the graph representation for the image to be segmented comprises:
establishing nodes,
each node corresponding to one voxel in the multi-parametric MR image and the weight of each edge connecting two nodes corresponds to the feature similarity measurement between the two nodes connected.
6. The method of claim 4 , wherein each vertex in the graph has a label and the method further comprises:
determining whether the SVM classification is reliable for each node based on the classification probability, and initializing the label information of the node if reliable.
7. The method of claim 5 , wherein said feature similarity measurement between connected nodes compromises:
the multi-parametric image intensity similarity, the spatial proximity, and edge information between verticies in the multi-parametric image being considered.
8. The method of claim 1 , wherein said step of generating the segmented brain tumor image comprises:
optimizing a corresponding object function for segmentation iteratively based on the initial label information and an edge weight in the graph.
9. A method for brain tumor segmentation in multi-parametric 3D magnetic resonance (MR) images, comprising:
pre-processing an input multi-parametric 3D MR image;
classifying each voxel in the pre-processed multi-parametric 3D MR image by a support vector machine (SVM) based classifier to determine a classification probability that the voxel is part of a brain tumor for each voxel;
labeling those voxels for which the determined classification probability exceeds a selected threshold as initial label information for image segmentation;
constructing a graph based representation for the pre-processed image to be segmented; and
generating the segmented brain tumor image using the initial label information and graph based representation.
10. The method of claim 9 , further comprising:
constructing a feature vector for each voxel based on the multi-parametric image intensity information of voxels in its spatial neighborhood.
11. The method of claim 10 , further comprising:
classifying the feature vector of each voxel using the SVM classification method.
12. The method of claim 9 , wherein constructing the graph based representation for the image to be segmented, further comprises:
establishing nodes, each node corresponding to one voxel in the multi-parametric MR image and the weight of each edge connecting two nodes corresponds to a feature similarity measurement between the two nodes connected.
13. The method of claim 9 , wherein each vertex in the graph based representation has a label and the method further comprises:
determining whether the SVM classification is reliable for each node based on the classification probability, and initializing the label information of the node if reliable.
14. The method of claim 13 , wherein said feature similarity measurement between connected nodes compromises the multi-parametric image intensity similarity, the spatial proximity, and edge information between verticies in the multi-parametric image being considered.
15. The method of claim 9 , wherein said step of generating the segmented brain tumor image comprises:
optimizing a corresponding object function for segmentation iteratively based on the initial label information and an edge weight in the graph.