Method and system for detection and registration of 3D objects using incremental parameter learning

• US 20080211812A1
• Filed: 02/01/2008
• Published: 09/04/2008
• Est. Priority Date: 02/02/2007
• Status: Active Grant

• Current Assignee: Siemens Healthcare GMBH (Siemens AG)
• Original Assignee: Siemens AG
• Inventors: Bogoni, Luca, Salganicoff, Marcos, Lu, Le, Barbu, Adrian, Comaniciu, Dorin
• Granted Patent: US 8,068,654 B2
• Time in Patent Office: Days
• US Class Current: 345/424
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06V 10/24 Aligning, centring, orientation detection or correction of the image
• CPC Class Codes: G06V 10/774 Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06V 20/64 Three-dimensional objects
• CPC Class Codes: G06V 2201/031 of internal organs

Abstract

A method and system for detecting 3D objects in images is disclosed. In particular, a method and system for Ileo-Cecal Valve detection in 3D computed tomography (CT) images using incremental parameter learning and ICV specific prior learning is disclosed. First, second, and third classifiers are sequentially trained to detect candidates for position, scale, and orientation parameters of a box that bounds an object in 3D image. In the training of each sequential classifier, new training samples are generated by scanning the object'"'"'s configuration parameters in the current learning projected subspace (position, scale, orientation), based on detected candidates resulting from the previous training step. This allows simultaneous detection and registration of a 3D object with full 9 degrees of freedom. ICV specific prior learning can be used to detect candidate voxels for an orifice of the ICV and to detect initial ICV box candidates using a constrained orientation alignment at each candidate voxel.

Claims

1. An incremental parameter learning method for Ileo-Cecal Valve (ICV) detection in 3D computed tomography (CT) volumes, comprising:

receiving at least one 3D training CT volume;

training a first classifier which generates a number of ICV position box candidates for said at least one 3D training CT volume from a set of initial ICV box candidates;

training a second classifier which generates a number of ICV position and scale box candidates for said at least one 3D training CT volume from said ICV position box candidates; and

training a third classifier which detects a position, scale, and orientation of a 3D box bounding the ICV in said at least one 3D training volume from said ICV position and scale box candidates.

2. The method of claim 1, further comprising:

training an orifice classifier which generates a number of orifice candidate voxels in said at least one 3D training CT volume; and

training an initial orientation classifier which generates said set of initial ICV box candidates from said orifice candidate voxels.

3. The method of claim 2, wherein said step of training an orifice classifier comprises:

scanning voxels in said at least one 3D training CT volume on a colon surface to generate positive and negative training sets based on a distance between each voxel and an annotated orifice location in said at least one 3D training CT volume;

training said orifice classifier based on said positive and negative training sets using steerable features trained by a probabilistic boosting tree; and

classifying said voxels using said orifice classifier to detect said orifice candidate voxels.

4. The method of claim 2, wherein said step of training an initial orientation classifier comprises:

aligning an orientation of a 3D box centered at each of said orifice candidate voxels with a gradient vector at that orifice candidate voxel, wherein a scale of each 3D box is fixed;

generating training samples by rotating the 3D box centered at each of said orifice candidate voxels;

generating positive and negative training sets based on a distance between each training sample and an annotated ICV box in said at least one 3D training CT volume;

training said initial orientation classifier based on said positive and negative training sets using steerable features trained by a probabilistic boosting tree; and

classifying the training samples using said initial orientation classifier to detect said initial ICV box candidates.

5. The method of claim 1, wherein said step of training a first classifier comprises:

generating training samples by shifting a center position of each of said initial ICV box candidates;

generating positive and negative training sets based on a distance between each training sample and an annotated ICV box in said at least one 3D training CT volume;

training said first classifier based on said positive and negative training sets; and

classifying the training samples using said first classifier to detect said ICV position box candidates.

6. The method of claim 1, wherein said step of training a second classifier comprises:

generating training samples by varying a scale of each of said ICV position box candidates;

generating positive and negative training sets based on a distance between each training sample and an annotated ICV box in said at least one 3D training CT volume;

training said second classifier based on said positive and negative training sets; and

classifying the training samples using said second classifier to detect said ICV position and scale box candidates.

7. The method of claim 1, wherein said step of training a third classifier comprises:

generating training samples by adding disturbances to an orientation of each of said ICV position and scale box candidates;

generating positive and negative training sets based on a distance between each third training sample and an annotated ICV box in said at least one 3D training CT volume; and

training said third classifier based on said positive and negative training sets.

8. The method of claim 1, further comprising:

sequentially evaluating an input 3D CT image using said first, second, and third classifiers to detect the ICV in the input 3D CT image.

9. The method of claim 1, wherein each of said first, second, and third classifiers is trained using steerable features calculated based on a sampling grid pattern.

10. The method of claim 9, wherein a probabilistic boosting tree is used to train said steerable features for each of said first, second, and third classifiers.

11. A method for Ileo-Cecal Valve (ICV) detection in an input 3D computed tomography (CT) volume, comprising:

detecting initial box candidates for the ICV based an ICV orifice in said input 3D CT volume; and

detecting a box bounding the ICV in said 3D CT volume by sequentially detecting possible locations, scales, and orientations of the box bounding the ICV using incremental parameter learning based on said initial box candidates.

12. The method of claim 11, wherein said step of detecting initial box candidates for the ICV comprises:

detecting a number of ICV orifice candidate voxels in said 3D CT volume using a trained 3D point detector;

aligning an orientation of a 3D box centered at each orifice candidate voxel with a gradient vector at that orifice candidate voxel, wherein a scale of each 3D box is fixed;

generating a set testing boxes by rotating the orientation of the 3D box centered at each orifice candidate voxel; and

detecting a number of said testing boxes as initial box candidates using a trained 3D box detector.

13. The method of claim 11, wherein said step of detecting a box bounding the ICV in said 3D CT volume comprises:

generating a plurality of first testing boxes by shifting a center location of each of said initial box candidates;

detecting a number of said first testing boxes as ICV position box candidates using a first trained classifier;

generating a plurality of second testing boxes by varying a scale of each of said ICV position box candidates;

detecting a number of said second testing boxes as ICV position and scale box candidates using a second trained classifier;

generating a plurality of third testing boxes by adding disturbances to an orientation of each of said ICV position and scale box candidates; and

detecting one of said third testing boxes as the box bounding the ICV in said 3D CT volume.

14. The method of claim 13, wherein each of said first, second, and third trained classifiers is trained using a probabilistic boosting tree using steerable features.

15. An apparatus for incremental parameter learning method for Ileo-Cecal Valve (ICV) detection in 3D computed tomography (CT) volumes, comprising:

means for receiving at least one 3D training CT volume;

means for training a first classifier which generates a number of ICV position box candidates for said at least one 3D training CT volume from a set of initial ICV box candidates;

means for training a second classifier which generates a number of ICV position and scale box candidates for said at least one 3D training Ct volume from said ICV position box candidates; and

means for training a third classifier which detects a position, scale, and orientation of a 3D box bounding the ICV in said at least one 3D training volume from said ICV position and scale box candidates.

16. The apparatus of claim 15, further comprising:

means for training an orifice classifier which generates a number of orifice candidate voxels in said at least one 3D training CT volume; and

means for training an initial orientation classifier which generates said set of initial ICV box candidates from said orifice candidate voxels.

17. The apparatus of claim 15, wherein said means for training a first classifier comprises:

means for generating training samples by shifting a center position of each of said initial ICV box candidates;

means for generating positive and negative training sets based on a distance between each training sample and an annotated ICV box in said at least one 3D training CT volume;

means for training said first classifier based on said positive and negative training sets; and

means for classifying the training samples using said first classifier to detect said ICV position box candidates.

18. The apparatus of claim 15, wherein said means for training a second classifier comprises:

means for generating training samples by varying a scale of each of said ICV position box candidates;

means for generating positive and negative training sets based on a distance between each training sample and an annotated ICV box in said at least one 3D training CT volume;

means for training said second classifier based on said positive and negative training sets; and

means for classifying the training samples using said second classifier to detect said ICV position and scale box candidates.

19. The apparatus of claim 15, wherein said means for training a third classifier comprises:

means for generating training samples by adding disturbances to an orientation of each of said ICV position and scale box candidates;

means for generating positive and negative training sets based on a distance between each third training sample and an annotated ICV box in said at least one 3D training CT volume; and

means for training said third classifier based on said positive and negative training sets.

20. An apparatus for Ileo-Cecal Valve (ICV) detection in an input 3D computed tomography (CT) volume, comprising:

means for detecting initial box candidates for the ICV based an ICV orifice in said input 3D CT volume; and

means for detecting a box bounding the ICV in said 3D CT volume by sequentially detecting possible locations, scales, and orientations of the box bounding the ICV using incremental parameter learning based on said initial box candidates.

21. The apparatus of claim 20, wherein said means for detecting initial box candidates for the ICV comprises:

means for detecting a number of ICV orifice candidate voxels in said 3D CT volume using a trained 3D point detector;

means for aligning an orientation of a 3D box centered at each orifice candidate voxel with a gradient vector at that orifice candidate voxel, wherein a scale of each 3D box is fixed;

means for generating a set testing boxes by rotating the orientation of the 3D box centered at each orifice candidate voxel; and

means for detecting a number of said testing boxes as initial box candidates using a trained 3D box detector.

22. The apparatus of claim 20, wherein said means for detecting a box bounding the ICV in said 3D CT volume comprises:

means for generating a plurality of first testing boxes by shifting a center location of each of said initial box candidates;

means for detecting a number of said first testing boxes as ICV position box candidates using a first trained classifier;

means for generating a plurality of second testing boxes by varying a scale of each of said ICV position box candidates;

means for detecting a number of said second testing boxes as ICV position and scale box candidates using a second trained classifier;

means for generating a plurality of third testing boxes by adding disturbances to an orientation of each of said ICV position and scale box candidates; and

means for detecting one of said third testing boxes as the box bounding the ICV in said 3D CT volume.

23. A computer readable medium encoded by computer executable instructions for incremental parameter learning for Ileo-Cecal Valve (ICV) detection in 3D computed tomography (CT) volumes, the computer executable instructions defining steps comprising:

receiving at least one 3D training CT volume;

training a first classifier which generates a number of ICV position box candidates for said at least one 3D training CT volume from a set of initial ICV box candidates;

training a second classifier which generates a number of ICV position and scale box candidates for said at least one 3D training Ct volume from said ICV position box candidates; and

training a third classifier which detects a position, scale, and orientation of a 3D box bounding the ICV in said at least one 3D training volume from said ICV position and scale box candidates.

24. The computer readable medium of claim 23, further comprising computer executable instructions defining the steps of:

training an orifice classifier which generates a number of orifice candidate voxels in said at least one 3D training CT volume; and

training an initial orientation classifier which generates said set of initial ICV box candidates from said orifice candidate voxels.

25. The computer readable medium of claim 23, further comprising computer executable instructions defining the step of:

sequentially evaluating an input 3D CT image using said first, second, and third classifiers to detect the ICV in the input 3D CT image