System and method for local pulmonary structure classification for computer-aided nodule detection

• US 20070172105A1
• Filed: 01/18/2007
• Published: 07/26/2007
• Est. Priority Date: 01/25/2006
• Status: Active Grant

• Current Assignee: Siemens Medical Solutions USA Incorporated (Siemens AG)
• Original Assignee: Siemens Medical Solutions USA Incorporated (Siemens AG)
• Inventors: Okada, Kazunori, Bahlmann, Claus, Li, Xianlin
• Granted Patent: US 7,764,819 B2
• Time in Patent Office: Days
• US Class Current: 382/131
• CPC Class Codes: A61B 2090/364 Correlation of different images or relation of image positions in respect to the body
• CPC Class Codes: G06F 18/22 Matching criteria, e.g. proximity measures
• CPC Class Codes: G06F 18/2321 using statistics or function optimisation, e.g. modelling of probability density functions
• CPC Class Codes: G06V 10/761 Proximity, similarity or dissimilarity measures
• CPC Class Codes: G06V 10/763 Non-hierarchical techniques, e.g. based on statistics of modelling distributions
• CPC Class Codes: G06V 2201/03 Recognition of patterns in medical or anatomical images

Abstract

A method for classifying pulmonary structures in digitized images includes providing approximate target structure locations of one or more target structures in a digitized 3-dimensional (3D) image, fitting an anisotropic Gaussian model about said approximate target locations to generate more precise 3D target models and center locations of said one or more target structures, warping each said 3D target model into a 3D sphere, constructing a bounding manifold about each said warped 3D sphere, and identifying clusters on said bounding manifold wherein said one or more target structures are classified.

Claims

1. A method for classifying pulmonary structures in digitized images, comprising the steps of:

providing approximate target structure locations of one or more target structures in a digitized 3-dimensional (3D) image;

fitting an anisotropic Gaussian model about said approximate target locations to generate more precise 3D target models and center locations of said one or more target structures;

warping each said 3D target models into a 3D sphere;

constructing a bounding manifold about each said warped 3D sphere; and

identifying clusters on said bounding manifolds wherein said one or more target structures are classified.

2. The method of claim 1, wherein said digitized image comprises a plurality of intensities corresponding to a domain of points on a 3-dimensional grid.

3. The method of claim 1, wherein fitting an anisotropic Gaussian model about an approximate target location comprises using Gaussian scale-space mean shift analysis and Jensen-Shannon divergence-based automatic bandwidth selection generating a 3D ellipsoidal model of said target structure, wherein the center and dimensions of said 3D ellipsoid correspond to the center and covariances of said Gaussian model.

4. The method of claim 3, wherein warping said 3D target model comprises affine-normalizing said 3D ellipsoid wherein scaling directions and factors are obtained from the structure covariance of said anisotropic Gaussian model.

5. The method of claim 1, wherein constructing a bounding manifold further comprises unwrapping the 3D surface of the warped sphere into a 2D representation, and determining a radius of an appropriate bounding manifold.

6. The method of claim 5, wherein unwrapping the 3D surface of the warped sphere into a 2D representation comprises transforming the surface of said warped sphere into spherical coordinates (θ

, φ

) wherein φ

ε

[−

π

,π

] and θ

ε

−

π

,π

].

7. The method of claim 5, wherein determining a radius of an appropriate bounding manifold comprises constructing a plurality of spherical manifolds of different radii about said warped sphere, unwrapping each spherical manifold into a 2D representation, normalizing the intensity value distribution on each said unwrapped spherical manifold, calculating an intensity entropy for each said unwrapped spherical manifold wherein intensity values are treated as probability values wherein an entropy distribution is defined, and finding a radius that minimizes said entropy distribution.

8. The method of claim 1, wherein identifying clusters comprises using an expectation-maximization to fit a mixture

9. The method of claim 1, further comprising using agglomerative hierarchical clustering to merge clusters within a predefined distance of each other, using a distance metric for a pair of multivariate wrapped Gaussian distributions equivalent to

10. The method of claim 1, wherein the pulmonary structure class is determined by the number of wrapped Gaussian component clusters associated with a target structure, wherein a solitary nodule has 0 clusters, an attached nodule has 2 clusters, a vessel has 4 clusters, and a vessel junction has 6 or more clusters.

11. A method for classifying pulmonary structures in digitized images, comprising the steps of:

providing target locations of one or more 3D spheres in a digitized 3-dimensional (3D) image, said image comprising a a plurality of intensities corresponding to a domain of points on a 3-dimensional grid, each 3D sphere representing a target structure in said image;

constructing a plurality of spherical manifolds of different radii about said 3D sphere;

calculating an intensity entropy for each said spherical manifold wherein intensity values are treated as probability values wherein an entropy distribution is defined;

finding a radius that minimizes said entropy distribution, wherein said minimizing radius defines a bounding manifold;

unwrapping the surface of bounding manifold into a 2D spherical coordinate (θ

, φ

) representation wherein φ

ε

[−

π

,π

] and θ

ε

[−

π

,π

];

using expectation-maximization to fit a mixture

$N_w<br><br>\left(\mathrm{\Theta <br><br>}\right)=\underset{p=1}{\overset{P}{\sum <br><br>}}<br><br>c_p<br><br>N_w^p<br><br>\left(\mathrm{\Theta <br><br>}\right)<br><br>\text{\hspace{1em}<br><br>}$

multivariate wrapped Gaussian distribution N_w(Θ

) of a vector variable Θ

=(₁, . . . , ^F)^T, wherein mixture component probabilities c_p are estimated within the expectation-maximization, wherein _i=(θ

_i,φ

_i) to clusters of target structures protruding through said bounding manifold, and wherein a pulmonary structure is classified by a number of protruding clusters.

12. The method of claim 11, further comprising normalizing the intensity distribution on each of said plurality of spherical manifolds.

13. The method of claim 11, wherein providing target locations of one or more 3D sphere comprises providing approximate target structure locations of said one or more target structures in said digitized 3-dimensional (3D) image;

fitting an anisotropic Gaussian model about said approximate target locations to generate more precise 3D ellipsoidal target models and center locations of said ellipsoidal models; and

affine-normalizing said ellipsoidal models into a 3D sphere.

14. The method of claim 11, wherein in each dimension _i satisfies =x_w=x mod 2π

ε

(−

π

,π

], N_w(Θ

) satisfies

15. A program storage device readable by a computer, tangibly embodying a program of instructions executable by the computer to perform the method steps for classifying pulmonary structures in digitized images, said method comprising the steps of:

providing approximate target structure locations of one or more target structures in a digitized 3-dimensional (3D) image;

fitting an anisotropic Gaussian model about said approximate target locations to generate more precise 3D target models and center locations of said one or more target structures;

warping each said 3D target model into a 3D sphere;

constructing a bounding manifold about each said warped 3D sphere; and

identifying clusters on said bounding manifold wherein said one or more target structures are classified.

16. The computer readable program storage device of claim 15, wherein said digitized image comprises a plurality of intensities corresponding to a domain of points on a 3-dimensional grid.

17. The computer readable program storage device of claim 15, wherein fitting an anisotropic Gaussian model about an approximate target location comprises using Gaussian scale-space mean shift analysis and Jensen-Shannon divergence-based automatic bandwidth selection generating a 3D ellipsoidal model of said target structure, wherein the center and dimensions of said 3D ellipsoid correspond to the center and covariances of said Gaussian model.

18. The computer readable program storage device of claim 17, wherein warping said 3D target model comprises affine-normalizing said 3D ellipsoid wherein scaling directions and factors are obtained from the structure covariance of said anisotropic Gaussian model.

19. The computer readable program storage device of claim 15, wherein constructing a bounding manifold further comprises unwrapping the 3D surface of the warped sphere into a 2D representation, and determining a radius of an appropriate bounding manifold.

20. The computer readable program storage device of claim 19, wherein unwrapping the 3D surface of the warped sphere into a 2D representation comprises transforming the surface of said warped sphere into spherical coordinates (θ

, φ

) wherein φ

ε

[−

π

,π

] and θ

ε

[−

π

,π

].

21. The computer readable program storage device of claim 19, wherein determining a radius of an appropriate bounding manifold comprises constructing a plurality of spherical manifolds of different radii about said warped sphere, unwrapping each spherical manifold into a 2D representation, normalizing the intensity value distribution on each said unwrapped spherical manifold, calculating an intensity entropy for each said unwrapped spherical manifold wherein intensity values are treated as probability values wherein an entropy distribution is defined, and finding a radius that minimizes said entropy distribution.

22. The computer readable program storage device of claim 15, wherein identifying clusters comprises using an expectation-maximization to fit a mixture

23. The computer readable program storage device of claim 15, the method further comprising using agglomerative hierarchical clustering to merge clusters within a predefined distance of each other, using a distance metric for a pair of multivariate wrapped Gaussian distributions equivalent to

24. The computer readable program storage device of claim 15, wherein the pulmonary structure class is determined by the number of wrapped Gaussian component clusters associated with a target structure, wherein a solitary nodule has 0 clusters, an attached nodule has 2 clusters, a vessel has 4 clusters, and a vessel junction has 6 or more clusters.