Method and system for automatic prostate segmentation in magnetic resonance images

1. A method for automatic prostate segmentation in magnetic resonance (MR) image data, comprising:

performing a first intensity normalization on the MR image to adjust for global contrast changes between the MR image and other MR scans;

performing a second intensity normalization on the MR image to adjust for intensity variation within the MR image due to an endorectal coil used to acquire the MR image by:

obtaining a mask image from the MR image using intensity thresholding,

extracting a bright region from the MR image using the mask image, and

calculating adjusted intensities to reduce an overall intensity within the bright region such that the adjusted intensities at a boundary of the bright region match a surrounding region in the MR image and gradient features within the bright region are retained;

obtaining an initial prostate segmentation in the MR image by aligning a learned statistical shape model of the prostate to the MR image using marginal space learning (MSL); and

refining the initial prostate segmentation using one or more trained boundary classifiers.

2. The method of claim 1 , wherein the MR image is a 3D MR image resulting from a T2-weighted axial abdominal MR scan.

3. The method of claim 1 , wherein performing the first intensity normalization on the MR image to adjust for global contrast changes between the MR image and other MR scans comprises:

calculating a linear transformation of the MR image that minimizes a least square error between at least a portion of intensity distributions of a stored target image and the MR image; and

adjusting intensities of the MR image using the linear transformation.

4. The method of claim 3 , wherein:

calculating the linear transformation of the MR image that minimizes the least square error between at least the portion of intensity distributions of the stored target image and the MR image comprises calculating the linear transformation as:

a

,

b

=

arg

⁢

⁢

min

a

,

b

⁢

∑

j

=

3

98

⁢

(

prctile

⁡

(

I

^

,

j

)

-

(

prctile

⁡

(

I

i

,

j

)

⁢

a

+

b

)

)

2

,

where Î is the stored target image, I i is the MR image, prctile(Î, j) is the j th percentile of the intensities of the stored target image and prctile(I i , j) is the j th percentile of the intensities of the MR image; and

adjusting intensities of the MR image using the linear transformation comprises generating an adjusted image as: I i ′=I i a+b.

5. The method of claim 1 , wherein performing the second intensity normalization on the MR image to adjust for intensity variation within the MR image due to the endorectal coil used to acquire the MR image comprises:

obtaining the mask image as M=((I>τ 1 )⊕B) (I>τ 2 ), where I is the MR image, τ 1 and τ 2 are intensity thresholds, τ 1 >τ 2 , and ⊕B is a dilation with a circular ball;

extracting the bright region Ω R 2 from the MR image Ω R 2 , as the non-zero elements of the mask image M;

generating a high pass filtered image as g(x)=(I−G σ *I)(x), where G σ is a Gaussian function; and

calculating adjusted intensities within the bright region f: Ω R as E(f)=min∫ Ω |∇f−∇g| 2 dx where f=I on δΩ by solving a Poisson equation: ∇ 2 f=∇ 2 g.

6. The method of claim 1 , wherein the statistical shape model of the prostate represents the shape of the prostate as a linear combination of a mean prostate shape and a number of strongest shape modes each weighted by a respective shape coefficient.

7. The method of claim 6 , wherein the mean prostate shape and the number of strongest shape modes are learned from a set of annotated training images.

8. The method of claim 6 , wherein obtaining the initial prostate segmentation in the MR image by aligning the learned statistical shape model of the prostate to the MR image using marginal space learning (MSL) comprises:

detecting position candidates for the prostate in the MR image using a first discriminative classifier;

detecting position-orientation candidates for the prostate in the MR image based on the detected position candidates using a second discriminative classifier;

detecting position-orientation-scale candidates for the prostate in the MR image based on the detected position-orientation candidates using a third discriminative classifier, wherein each position-orientation-scale candidate defines a candidate bounding box for aligning the mean prostate shape to the MR image; and

detecting the respective shape coefficient for each of the number of strongest shape modes based on the detected position-orientation-scale candidates using a fourth discriminative classifier.

9. The method of claim 8 , wherein the number of strongest shape modes comprises three strongest shape modes.

10. The method of claim 1 , wherein refining the initial prostate segmentation using one or more trained boundary classifiers comprises:

refining a prostate surface mesh resulting from the initial prostate segmentation by calculating a displacement for each of a plurality of vertices of the mesh along a surface normal using the one or more trained boundary classifiers; and

projecting a shape of the refined prostate surface mesh to a learned statistical shape space of the prostate.

11. The method of claim 1 , wherein refining the initial prostate segmentation using one or more trained boundary classifiers comprises:

refining a low resolution prostate surface mesh resulting from the initial prostate segmentation by performing a number of iterations of:

refining vertices of the low resolution prostate surface mesh using a first trained boundary classifier, and

projecting a shape of the refined low resolution prostate surface mesh to a learned statistical shape space of the prostate;

upsampling the refined low resolution prostate surface mesh to generate a high resolution prostate surface mesh; and

refining the high resolution prostate surface mesh by performing a number of iterations of:

refining vertices of the high resolution prostate surface mesh using one or more second trained boundary classifiers, and

projecting a shape of the refined high resolution prostate surface mesh to the learned statistical shape space of the prostate.

12. The method of claim 11 , wherein refining vertices of the high resolution prostate surface mesh using one or more second trained boundary classifiers comprises:

partitioning the high resolution prostate surface mesh into a plurality of regions; and

refining the vertices of the high resolution surface mesh using a different discriminative classifier for each of the plurality of regions.

13. An apparatus for automatic prostate segmentation in magnetic resonance (MR) image data, comprising:

means for performing a first intensity normalization on the MR image to adjust for global contrast changes between the MR image and other MR scans;

means for performing a second intensity normalization on the MR image to adjust for intensity variation within the MR image due to an endorectal coil used to acquire the MR image, comprising;

means for obtaining a mask image from the MR image;

means for extracting a bright region from the MR image using the mask image, and

means for calculating adjusted intensities to reduce an overall intensity within the bright region such that the adjusted intensities at a boundary of the bright region match a surrounding region in the MR image and gradient features within the bright region are retained;

means for obtaining an initial prostate segmentation in the MR image by aligning a learned statistical shape model of the prostate to the MR image using marginal space learning (MSL); and

means for refining the initial prostate segmentation using one or more trained boundary classifiers.

14. The apparatus of claim 13 , wherein the statistical shape model of the prostate represents the shape of the prostate as a linear combination of a mean prostate shape and a number of strongest shape modes each weighted by a respective shape coefficient.

15. The apparatus of claim 14 , wherein the means for obtaining the initial prostate segmentation in the MR image by aligning the learned statistical shape model of the prostate to the MR image using marginal space learning (MSL) comprises:

means for detecting position candidates for the prostate in the MR image using a first discriminative classifier;

means for detecting position-orientation candidates for the prostate in the MR image based on the detected position candidates using a second discriminative classifier;

means for detecting position-orientation-scale candidates for the prostate in the MR image based on the detected position-orientation candidates using a third discriminative classifier, wherein each position-orientation-scale candidate defines a candidate bounding box for aligning the mean prostate shape to the MR image; and

means for detecting the respective shape coefficient for each of the number of strongest shape modes based on the detected position-orientation-scale candidates using a fourth discriminative classifier.

16. The apparatus of claim 13 , wherein the means for refining the initial prostate segmentation using one or more trained boundary classifiers comprises:

means for refining a prostate surface mesh resulting from the initial prostate segmentation by calculating a displacement for each of a plurality of vertices of the mesh along a surface normal using the one or more trained boundary classifiers; and

means for projecting a shape of the refined prostate surface mesh to a learned statistical shape space of the prostate.

17. A non-transitory computer readable medium storing computer program instructions for automatic prostate segmentation in magnetic resonance (MR) image data, the computer program instructions when executed on a processor cause the processor to perform operations comprising:

performing a first intensity normalization on the MR image to adjust for global contrast changes between the MR image and other MR scans;

performing a second intensity normalization on the MR image to adjust for intensity variation within the MR image due to an endorectal coil used to acquire the MR image by:

obtaining a mask image from the MR image using intensity thresholding,

extracting a bright region from the MR image using the mask image, and

calculating adjusted intensities to reduce an overall intensity within the bright region such that the adjusted intensities at a boundary of the bright region match a surrounding region in the MR image and gradient features within the bright region are retained;

obtaining an initial prostate segmentation in the MR image by aligning a learned statistical shape model of the prostate to the MR image using marginal space learning (MSL); and

refining the initial prostate segmentation using one or more trained boundary classifiers.

18. The non-transitory computer readable medium of claim 17 , wherein performing the first intensity normalization on the MR image to adjust for global contrast changes between the MR image and other MR scans comprises:

calculating a linear transformation of the MR image that minimizes a least square error between at least a portion of intensity distributions of a stored target image and the MR image; and

adjusting intensities of the MR image using the linear transformation.

19. The non-transitory computer readable medium of claim 17 , wherein the statistical shape model of the prostate represents the shape of the prostate as a linear combination of a mean prostate shape and a number of strongest shape modes each weighted by a respective shape coefficient.

20. The non-transitory computer readable medium of claim 19 , wherein obtaining the initial prostate segmentation in the MR image by aligning the learned statistical shape model of the prostate to the MR image using marginal space learning (MSL) comprises:

detecting position candidates for the prostate in the MR image using a first discriminative classifier;

detecting position-orientation candidates for the prostate in the MR image based on the detected position candidates using a second discriminative classifier;

detecting position-orientation-scale candidates for the prostate in the MR image based on the detected position-orientation candidates using a third discriminative classifier, wherein each position-orientation-scale candidate defines a candidate bounding box for aligning the mean prostate shape to the MR image; and

detecting the respective shape coefficient for each of the number of strongest shape modes based on the detected position-orientation-scale candidates using a fourth discriminative classifier.

21. The non-transitory computer readable medium of claim 17 , wherein refining the initial prostate segmentation using one or more trained boundary classifiers comprises:

refining a prostate surface mesh resulting from the initial prostate segmentation by calculating a displacement for each of a plurality of vertices of the mesh along a surface normal using the one or more trained boundary classifiers; and

projecting a shape of the refined prostate surface mesh to a learned statistical shape space of the prostate.