Image-based detection using hierarchical learning

1. A method of detecting anatomical components in images, comprising:

(i) receiving a medical image;

(ii) detecting at least one anchor landmark in the image by applying an anchor landmark detector that is learned from a first set of training samples;

(iii) detecting, by a processor, at least one bundle landmark in the image based at least in part on a position of the anchor landmark, wherein the bundle landmark is detected by applying a bundle landmark detector that is learned from a second set of training samples including positive and negative samples, wherein the positive samples include voxels close to any bundle landmark within a bundle of anatomical structures that are grouped based on similar characteristics and are associated with the bundle landmark detector, wherein the negative samples include remaining voxels in a local volume containing landmarks of one or more neighboring bundles; and

(iv) detecting, by the processor, at least one anatomical primitive based at least in part on at least two neighboring landmarks by applying an anatomical primitive detector that is learned from a third set of training samples, wherein the third set of training samples includes positive samples extracted by sampling voxels located within a specific anatomical primitive and a voxel within the specific anatomical primitive is treated as an independent positive sample.

2. The method of claim 1 wherein the anchor landmark and the bundle landmark represent vertebrae.

3. The method of claim 1 wherein the anchor landmark and the bundle landmark represent inter-vertebral discs.

4. The method of claim 1 further comprising:

aligning a set of images; and

extracting the second set of training samples from the aligned images.

5. The method of claim 4 further comprising:

extracting, from the aligned images, the positive samples by sampling voxels close to any bundle landmark within the bundle associated with the bundle landmark detector; and

extracting, from the aligned images, the negative samples by sampling the remaining voxels in the local volume containing the landmarks of the one or more neighboring bundles.

6. The method of claim 1 wherein detecting the bundle landmark comprises:

determining a first region of search based on the anchor landmark; and

applying the bundle landmark detector within the first region of search to detect the bundle landmark.

7. The method of claim 6 wherein determining the region of search comprises:

adapting a first local articulation model based at least in part on the anchor landmark; and

determining the first region of search using the adapted first local articulation model.

8. The method of claim 7 further comprising validating and labeling the detected bundle landmark using the first local articulation model.

9. The method of claim 1 wherein the anatomical primitive represents an inter-vertebral disc.

10. The method of claim 1 wherein detecting the anatomical primitive comprises:

determining a second region of search based on the neighboring landmarks; and

applying an anatomical primitive detector within the second region of search.

11. The method of claim 10 wherein determining the second region of search comprises:

adapting a second local articulation model based at least in part on the neighboring landmarks; and

determining the second region of search using the adapted second local articulation model.

12. The method of claim 11 further comprising validating and labeling the detected anatomical primitive using the second local articulation model.

13. The method of claim 1 wherein the third set of training samples includes negative samples obtained by sampling remaining voxels in a local volume containing neighboring structures.

14. The method of claim 13 wherein the anchor landmark is detected by applying an anchor landmark detector that is learned from a first set of training samples.

15. The method of claim 13 wherein the bundle landmark is detected by applying a bundle landmark detector that is learned from a second set of training samples.

16. A method of detecting anatomical components in images, comprising:

(i) receiving a medical image;

(ii) detecting at least two neighboring landmarks grouped in a same bundle with similar characteristics in the image; and

(iii) determining a region of search using a local articulation model adapted based at least in part on the neighboring landmarks; and

(iv) detecting, by a processor, at least one anatomical primitive in the image by applying an anatomical primitive detector within the region of search by applying an anatomical primitive detector that is learned from a set of training samples that includes positive samples extracted by sampling voxels located within a specific anatomical primitive and treating a voxel within the specific anatomical primitive as an independent positive sample.

17. A non-transitory computer readable medium embodying a program of instructions executable by machine to perform steps for detecting anatomical components in images, the steps comprising:

(i) receiving a medical image;

(ii) detecting at least one anchor landmark in the image by applying an anchor landmark detector that is learned from a first set of training samples;

(iii) detecting at least one bundle landmark in the image based at least in part on a position of the anchor landmark, wherein the bundle landmark is detected by applying a bundle landmark detector that is learned from a second set of training samples; and

(iv) detecting at least one anatomical primitive based at least in part on at least two neighboring landmarks by applying an anatomical primitive detector that is learned from a third set of training samples, wherein the third set of training samples includes positive samples and negative samples, wherein the positive samples are extracted by sampling voxels located within a specific anatomical primitive and a voxel within the specific anatomical primitive is treated as an independent positive sample, wherein the negative samples are obtained by sampling remaining voxels in a local volume containing neighboring structures.

18. A system for detecting anatomical components in images, comprising:

a non-transitory memory device for storing computer readable program code;

a processor in communication with the memory device, the processor being operative with the computer readable program code to:

(i) receive a medical image;

(ii) detect at least one anchor landmark in the image by applying an anchor landmark detector that is learned from a first set of training samples; and

(iii) detect at least one bundle landmark in the image based at least in part on a position of the anchor landmark, wherein the bundle landmark is detected by applying a bundle landmark detector that is learned from a second set of training samples; and

(iv) detecting at least one anatomical primitive based at least in part on at least two neighboring landmarks by applying an anatomical primitive detector that is learned from a third set of training samples, wherein the positive samples are extracted by sampling voxels located within a specific anatomical primitive and a voxel within the specific anatomical primitive is treated as an independent positive sample, wherein the negative samples are obtained by sampling remaining voxels in a local volume containing neighboring structures.