Automatic segmentation of articulated structures

1. A non-transitory computer-readable medium embodying a program of instructions executable by machine to perform steps for segmenting articulated structures, the steps comprising:

(i) generating statistical shape models and local appearance models based on a reference image and training images;

(ii) receiving a target image;

(iii) determining a non-rigid transformation function that registers first centerline landmarks along centerlines of articulated structures in the reference image with second centerline landmarks along centerlines of articulated structures in the target image;

(iv) deforming mean shapes of the statistical shape models to the target image space by applying the non-rigid transformation function on the mean shapes;

(v) decoupling any overlapping segments of articulated structures in the mean shapes;

(vi) searching for candidate points in the mean shapes using the local appearance models; and

(vii) fitting the mean shapes to the candidate points to generate a segmentation mask.

2. The computer-readable medium of claim 1 embodying further instructions executable by the machine to automatically detect the second centerline landmarks by using a learned landmark detector.

3. A computer-implemented method of segmenting articulated structures, comprising:

(i) receiving a target image, a reference image, statistical shape models, local appearance models and a learned landmark detector;

(ii) automatically detecting first centerline landmarks along centerlines of articulated structures in the target image using the learned landmark detector;

(iii) determining a non-rigid transformation function that registers second centerline landmarks along centerlines of articulated structures in the reference image with the first centerline landmarks;

(iv) deforming mean shapes of the statistical shape models to the target image space by applying the non-rigid transformation function on the mean shapes;

(v) searching for candidate points in the mean shapes using the local appearance models; and

(vi) fitting the mean shapes to the candidate points to generate a segmentation mask.

4. The method of claim 3 wherein automatically detecting the first centerline landmarks comprises:

invoking the learned landmark detector to automatically detect a point on at least one of the centerlines in the target image;

performing region growing by using the detected point as a seed to generate a mask;

extracting a skeleton of the mask; and

sampling points along the extracted skeleton to generate the first centerline landmarks.

5. The method of claim 3 wherein registering the second centerline landmarks with the first centerline landmarks comprises performing a thin-plate splines transformation function.

6. The method of claim 3 further comprising decoupling any overlapping segments of articulated structures in the mean shapes.

7. The method of claim 6 wherein decoupling any overlapping segments of the articulated structures comprises:

identifying a surface patch in a given first articulated structure that is intersecting with a neighboring second articulated structure;

determining a normal of the surface patch; and

iteratively moving apart the first and second articulated structures until the first and second articulated structures are separated.

8. The method of claim 3 further comprising repeating at least steps (v) and (vi) until a convergence criteria is satisfied.

9. The method of claim 3 further comprising generating the statistical shape models based on the reference image and training images.

10. The method of claim 9 wherein generating the statistical shape models comprise:

segmenting articulated structures in the reference image and the training images into disjoint segments;

extracting reference surfaces and first training surfaces of the disjoint segments in the reference image and the training images;

generating second training surfaces by deforming the reference surfaces to corresponding first training surfaces;

rigidly registering the second training surfaces to the reference surfaces; and

building the statistical shape models based on the registered second training surfaces.

11. The method of claim 10 wherein generating the second training surfaces comprises applying a deformation field that provides one-to-one point correspondences to the reference surfaces.

12. The method of claim 10 wherein rigidly registering the second training surfaces comprises applying a 7 degrees-of-freedom transformation to the second training surfaces.

13. The method of claim 10 wherein building the statistical shape models comprises generating a mean shape and eigenvectors of a covariance matrix of the registered second training surfaces by using eigenanalysis.

14. The method of claim 10 further comprising generating the local appearance models for points in the statistical shape models.

15. The method of claim 14 wherein generating the local appearance models comprises:

receiving a pair of reference surface and reference image and corresponding pairs of second training surface and training image associated with at least one of the statistical shape models;

dividing the reference surface and the second training surfaces into disjoint patches; and

generating local appearance models for the disjoint patches.

16. The method of claim 15 wherein at least one of the disjoint patches represent first 2-ring neighboring vertices around a given central vertex.

17. The method of claim 15 wherein generating the local appearance models comprises:

extracting appearance profiles from the reference and training images along surface normals of the disjoint patches; and

averaging the extracted appearance profiles to generate the local appearance models.

18. The method of claim 17 wherein extracting the appearance profiles comprises extracting the appearance profiles at different step sizes along the surface normals for different levels of scale.

19. An image segmentation system, comprising:

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

a processor in communication with the memory device, the processor being operative with the computer-readable program code to perform steps comprising

(i) receiving a target image, a reference image, statistical shape models and local appearance models,

(ii) determining a non-rigid transformation function that registers first centerline landmarks along centerlines of articulated structures in the reference image with second centerline landmarks along centerlines of articulated structures in the target image,

(iii) deforming mean shapes of the statistical shape models to the target image space by applying the non-rigid transformation function on the mean shapes,

(iv) searching for candidate points in the mean shapes using the local appearance models, and

(v) fitting the mean shapes to the candidate points to generate a segmentation mask.

20. The system of claim 19 wherein the articulated structures in the reference image and the target image comprise bone structures.