System and method for organ segmentation using surface patch classification in 2D and 3D images

1. A method for segmenting organs in digitized medical images, the method comprising:

using a computer to perform steps comprising:

providing a set of training images of an object of interest, wherein said object of interest has been segmented;

computing a surface mesh having a plurality of mesh cells that approximates a border of said object of interest;

extracting positive examples of all mesh cells and extracting negative examples in the neighborhood of each mesh cell which do not belong to the object surface;

training from said positive examples and negative examples a plurality of classifiers for outputting a probability of a point being a center of a particular mesh cell;

computing an active shape model using a robust subset of center points in said mesh cells;

wherein said robust subset of said mesh cells is determined by maximizing

arg

⁢

max

w

⁡

(

α

⁢

∑

i

n

⁢

⁢

∑

j

m

⁢

⁢

w

i

⁢

g

ij

+

β

⁢

∑

i

n

⁢

⁢

w

i

⁢

C

_

i

)

,

 g ij <D, wherein g ij is a geodesic distance between centers of mesh cells i and j, D is a maximum distance between mesh cells, C i is an average classification accuracy of said classifier for mesh cell i, α and β are adjustable parameters that can trade-off between sparsity and classification accuracy, and w i is a resulting weight of mesh cell i;

generating a new shape by iteratively deforming said active shape model to fit a test image; and

using said classifiers to calculate a probability of each center point of said new shape being a center of a particular mesh cell which said classifier was trained to recognize.

2. The method of claim 1 , further comprising aligning said images.

3. The method of claim 1 , further comprising computing a bounding box about said object of interest that is a smallest box that contains all points belonging to the object of interest, wherein said surface mesh is computed within the bounding box.

4. The method of claim 1 , wherein a number of mesh nodes of said surface mesh is constant for all sample images.

5. The method of claim 1 , wherein said object of interest is an organ.

6. The method of claim 1 , wherein said classifier is trained using positive and negative examples from said robust subset of mesh cells and their neighborhood.

7. The method of claim 1 , further comprising aligning said active shape model with an object of interest in a new image.

8. The method of claim 1 , wherein extracting positive and negative examples from said surface mesh comprises finding matching mesh cells on all surfaces in said training set for each mesh cell, selecting a set of center points in corresponding mesh cells in all images as positive examples, and selecting high gradient points that do not belong to the organ border within a predefined neighborhood of each mesh cell as negative examples.

9. A program storage device readable by a computer, tangibly embodying a program of instructions executable by the computer to perform the method steps for segmenting organs in digitized medical images, the method comprising the steps of:

providing a set of training images of an object of interest, wherein said object of interest has been segmented;

computing a surface mesh having a plurality of mesh cells that approximates a border of said object of interest;

extracting positive examples of all mesh cells and extracting negative examples in the neighborhood of each mesh cell which do not belong to the object surface;

training from said positive examples and negative examples a plurality of classifiers for outputting a probability of a point being a center of a particular mesh cell;

computing an active shape model using a robust subset of center points in said mesh cells;

wherein said robust subset of said mesh cells is determined by maximizing

arg

⁢

max

w

⁡

(

α

⁢

∑

i

n

⁢

⁢

∑

j

m

⁢

⁢

w

i

⁢

g

ij

+

β

⁢

∑

i

n

⁢

⁢

w

i

⁢

C

_

i

)

,

 g ij <D, wherein g ij is a geodesic distance between centers of mesh cells i and j, D is a maximum distance between mesh cells, C i is an average classification accuracy of said classifier for mesh cell i, α and β are adjustable parameters that can trade-off between sparsity and classification accuracy, and w i is a resulting weight of mesh cell i;

generating a new shape by iteratively deforming said active shape model to fit a test image; and

using said classifiers to calculate a probability of each center point of said new shape being a center of a particular mesh cell which said classifier was trained to recognize.

10. The computer readable program storage device of claim 9 , the method further comprising aligning said images.

11. The computer readable program storage device of claim 9 , the method further comprising computing a bounding box about said object of interest that is a smallest box that contains all points belonging to the object of interest, wherein said surface mesh is computed within the bounding box.

12. The computer readable program storage device of claim 9 , wherein a number of mesh nodes of said surface mesh is constant for all sample images.

13. The computer readable program storage device of claim 9 , wherein said object of interest is an organ.

14. The computer readable program storage device of claim 9 , wherein said classifier is trained using positive and negative examples from said robust subset of mesh cells and their neighborhood.

15. The computer readable program storage device of claim 9 , the method further comprising aligning said active shape model with an object of interest in a new image.

16. The computer readable program storage device of claim 9 , wherein extracting positive and negative examples from said surface mesh comprises finding matching mesh cells on all surfaces in said training set for each mesh cell, selecting a set of center points in corresponding mesh cells in all images as positive examples, and selecting high gradient points that do not belong to the organ border within a predefined neighborhood of each mesh cell as negative examples.