Method and system for brain tumor segmentation in 3D magnetic resonance images

1. A method for brain tumor segmentation in a multi-spectral 3D magnetic resonance (MR) image sequence comprising a plurality of voxels, comprising:

determining, for each voxel in the multi-spectral 3D MR image sequence, a probability that the voxel is part of a brain tumor, using a trained probabilistic boosting tree (PBT) classifier; and

generating a segmented brain tumor image using graph cuts segmentation based on the probabilities determined using the trained PBT classifier and intensities of the voxels in the multi-spectral 3D MR image sequence.

2. The method of claim 1 , further comprising:

pre-processing the multi-spectral 3D MR image prior to said determining step by:

performing skull stripping to remove skull image data in order to extract the brain in 3D MR images;

performing gradient anisotropic diffusion filtering; and

performing inter-scan intensity standardization.

3. The method of claim 1 , wherein said step of determining, for each voxel in the multi-spectral 3D MR image sequence, a probability that the voxel is part of a brain tumor, using a trained probabilistic boosting tree (PBT) classifier, comprises, for each voxel:

calculating local features for the voxel in a neighborhood of the voxel; and

determining the probability that the voxel is a brain tumor based on the local features calculated for the voxel.

4. The method of claim 3 , wherein said local features comprise at least one of individual multi-spectral intensities, inter-spectrality intensity gradients, 2D Haar-like features calculated on an intra-axial 2D context surrounding a voxel, and 3D Haar-like features calculated on a 3D context surrounding a voxel.

5. The method of claim 1 , wherein said step of generating a segmented brain tumor image using graph cuts segmentation based on the probabilities determined using the trained PBT classifier and intensities of the voxels in the multi-spectral 3D MR image sequence comprises:

optimizing an objective function that integrates the probabilities determined using the trained PBT classifier and the intensities of the voxels in the multi-spectral 3D MR image sequence.

6. The method of claim 5 , wherein the objective function comprises an observation model based on the probabilities determined using the trained PBT classifier and a spatial prior based on pair-wise homogeneity in classification labels and multi-spectral voxel intensities.

7. The method of claim 1 , wherein said step of generating a segmented brain tumor image using graph cuts segmentation based on the probabilities determined using the trained PBT classifier and intensities of the voxels in the multi-spectral 3D MR image sequence comprises:

constraining a graph cuts optimization based on the probabilities determined using the trained PBT classifier.

8. The method of claim 1 , wherein said step of generating a segmented brain tumor image using graph cuts segmentation based on the probabilities determined using the trained PBT classifier and intensities of the voxels in the multi-spectral 3D MR image sequence comprises:

generating a binary mask by optimizing an objective function based on the probabilities determined using the trained PBT classifier and intensities of the voxels in the multi-spectral 3D MR image sequence; and

generating the segmented brain tumor image by overlaying the binary mask on a 3D MR image of the multi-spectral 3D MR image sequence.

9. The method of claim 1 , further comprising:

removing false positive voxels from the segmented brain tumor image by retaining only a largest connected component of the segmented brain tumor image.

10. An apparatus for brain tumor segmentation in a multi-spectral 3D magnetic resonance (MR) image sequence comprising a plurality of voxels, comprising:

means for determining, for each voxel in the multi-spectral 3D MR image sequence, a probability that the voxel is part of a brain tumor, using a trained probabilistic boosting tree (PBT) classifier; and

means for generating a segmented brain tumor image using graph cuts segmentation based on the probabilities determined using the trained PBT classifier and intensities of the voxels in the multi-spectral 3D MR image sequence.

11. The apparatus of claim 10 , wherein said means for determining, for each voxel in the multi-spectral 3D MR image sequence, a probability that the voxel is a part of brain tumor, using a trained probabilistic boosting tree (PBT) classifier, comprises:

means for calculating local features for a voxel in a neighborhood of the voxel; and

means for determining the probability that the voxel is a brain tumor based on the local features calculated for the voxel.

12. The apparatus of claim 11 , wherein said local features comprise at least one of individual multi-spectral intensities, inter-spectrality intensity gradients, 2D Haar-like features calculated on an intra-axial 2D context surrounding a voxel, and 3D Haar-like features calculated on a 3D context surrounding a voxel.

13. The apparatus of claim 10 , wherein said means for generating a segmented brain tumor image using graph cuts segmentation based on the probabilities determined using the trained PBT classifier and intensities of the voxels in the multi-spectral 3D MR image sequence comprises:

means for optimizing an objective function that integrates the probabilities determined using the trained PBT classifier and the intensities of the voxels in the multi-spectral 3D MR image sequence.

14. The apparatus of claim 13 , wherein the objective function comprises an observation model based on the probabilities determined using the trained PBT classifier and a spatial prior based on pair-wise homogeneity in classification labels and multi-spectral voxel intensities.

15. The apparatus of claim 10 , wherein said means for generating a segmented brain tumor image using graph cuts segmentation based on the probabilities determined using the trained PBT classifier and intensities of the voxels in the multi-spectral 3D MR image sequence comprises:

means for constraining a graph cuts optimization based on the probabilities determined using the trained PBT classifier.

16. The apparatus of claim 10 , wherein said means for generating a segmented brain tumor image using graph cuts segmentation based on the probabilities determined using the trained PBT classifier and intensities of the voxels in the multi-spectral 3D MR image sequence comprises:

means for generating a binary mask by optimizing an objective function based on the probabilities determined using the trained PBT classifier and intensities of the voxels in the multi-spectral 3D MR image sequence; and

means for generating the segmented brain tumor image by overlaying the binary mask on a 3D MR image of the multi-spectral 3D MR image sequence.

17. A non-transitory computer readable medium encoded with computer executable instructions for brain tumor segmentation in a multi-spectral 3D magnetic resonance (MR) image sequence comprising a plurality of voxels, the computer executable instructions defining steps comprising:

determining, for each voxel in the multi-spectral 3D MR image sequence, a probability that the voxel is part of a brain tumor, using a trained probabilistic boosting tree (PBT) classifier; and

generating a segmented brain tumor image using graph cuts segmentation based on the probabilities determined using the trained PBT classifier and intensities of the voxels in the multi-spectral 3D MR image sequence.

18. The non-transitory computer readable medium of claim 17 , wherein the computer executable instructions defining the step of determining, for each voxel in the multi-spectral 3D MR image sequence, a probability that the voxel is part of a brain tumor, using a trained probabilistic boosting tree (PBT) classifier, comprise computer executable instructions defining the steps of, for each voxel:

calculating local features for the voxel in a neighborhood of the voxel; and

determining the probability that the voxel is a brain tumor based on the local features calculated for the voxel.

19. The non-transitory computer readable medium of claim 18 , wherein said local features comprise at least one of individual multi-spectral intensities, inter-spectrality intensity gradients, and 2D Haar-like features calculated on an intra-axial 2D context surrounding a voxel, and 3D Haar-like features calculated on a 3D context surrounding a voxel.

20. The non-transitory computer readable medium of claim 17 , wherein the computer executable instructions defining the step of generating a segmented brain tumor image using graph cuts segmentation based on the probabilities determined using the trained PBT classifier and intensities of the voxels in the multi-spectral 3D MR image sequence comprises computer executable instructions defining the step of:

optimizing an objective function that integrates the probabilities determined using the trained PBT classifier and the intensities of the voxels in the multi-spectral 3D MR image sequence.

21. The non-transitory computer readable medium of claim 20 , wherein the objective function comprises an observation model based on the probabilities determined using the trained PBT classifier and a spatial prior based on pair-wise homogeneity in classification labels and multi-spectral voxel intensities.

22. The non-transitory computer readable medium of claim 17 , wherein the computer executable instructions defining the step of generating a segmented brain tumor image using graph cuts segmentation based on the probabilities determined using the trained PBT classifier and intensities of the voxels in the multi-spectral 3D MR image sequence comprises computer executable instructions defining the step of:

constraining a graph cuts optimization based on the probabilities determined using the trained PBT classifier.

23. The non-transitory computer readable medium of claim 17 , wherein the computer executable instructions defining the step of generating a segmented brain tumor image using graph cuts segmentation based on the probabilities determined using the trained PBT classifier and intensities of the voxels in the multi-spectral 3D MR image sequence comprises computer executable instructions defining the steps of:

generating a binary mask by optimizing an objective function based on the probabilities determined using the trained PBT classifier and intensities of the voxels in the multi-spectral 3D MR image sequence; and

generating the segmented brain tumor image by overlaying the binary mask on a 3D MR image of the multi-spectral 3D MR image sequence.