Method for knowledge based image segmentation using shape models

• US 7,680,312 B2
• Filed: 05/08/2006
• Issued: 03/16/2010
• Est. Priority Date: 07/13/2005
• Status: Active Grant

• Current Assignee: Siemens Medical Solutions USA Incorporated (Siemens AG)
• Original Assignee: Siemens Medical Solutions USA Incorporated (Siemens AG)
• Inventors: Taron, Maxime G., Paragios, Nikolaos, Jolly, Marie-Pierre
• Primary Examiner(s): CHANG, JON CARLTON
• Application Number:
  US11/429,685
• Publication Number:
  US 20070014457A1
• Time in Patent Office: 1,408 Days
• Field of Search: 382/128, 382/131, 382/132, 382/173, 382/203
• US Class Current: 382/128
• CPC Class Codes: G06T 2207/10088 Magnetic resonance imaging [MRI]
• CPC Class Codes: G06T 2207/20081 Training; Learning
• CPC Class Codes: G06T 2207/30016 Brain
• CPC Class Codes: G06T 7/12 Edge-based segmentation
• CPC Class Codes: G06T 7/149 involving deformable models, e.g. active contour models
• CPC Class Codes: G06T 7/155 involving morphological operators

Abstract

A method for segmenting an object of interest from an image of a patient having such object. Each one of a plurality of training shapes is distorted to overlay a reference shape with a parameter Θ_i being a measure of the amount of distortion required to effect the overlay. A vector of the parameters Θ_i is obtained for every one of the training shapes through the minimization of a cost function along with an estimate of uncertainty for every one of the obtained vectors of parameters Θ_i, such uncertainty being quantified as a covariance matrix Σ_i. A statistical model represented as {circumflex over (f)}_H (Θ,Σ) is generated with the sum of kernels having a mean Θ_i and covariance Σ_i. The desired object of interest in the image of the patient is identified by positioning of the reference shape on the image and distorting the reference shape to overlay the obtained image with a parameter Θ being a measure of the amount of distortion required to effect the overlay. An uncertainty is quantified as a covariance matrix Σ and an energy function E=E_shape+E_image is computed to obtain the probability of the current shape in the statistical shape model E_shape(Θ,Σ)=−log({circumflex over (f)}_H) and the fit in the image E_image.

Claims

1. A method for use in segmenting an object of interest from an image of a patient having such object, comprising:

using medical imaging apparatus to perform the steps of;

transforming a generated reference shape of the object to match every one of a plurality of training shapes according to an energy function comprising distorting each one of the training shapes to overlay the reference shape with a parameter Θ

_i, being a measure of the amount of distortion required at each grid point to effect the overlay for the i^th one of the N training shapes;

obtaining a vector of the parameters Θ

_i, for every one of the training shapes through the minimization of a cost function;

estimating an uncertainty for every one of the obtained vectors of parameters Θ

_i, such uncertainty being quantified as a covariance matrix Σ

_i;

providing for the plurality of training shapes a statistical model represented as {circumflex over (f)}_H(Θ

,Σ

) which is the sum of the K Gaussian kernels having a mean Θ

_iand covariance Σ

_i.

2. The method recited in claim 1 wherein the cost function is the sum squared difference between a distance map generated for one of the training shapes and the transformed reference model.

3. The method recited in claim 1 including additionally:

obtaining the image of the desired object from the patient;

identifying the desired object of interest in the image of the patient by positioning of the reference shape on the image comprising;

transforming the generated reference shape to match the obtained image according to the energy function comprising distorting the reference shape to overlay the obtained image with a parameter Θ

being a measure of the amount of distortion required to effect the overlay;

estimating an uncertainty, such uncertainty being quantified as a covariance matrix Σ

;

computing a shape term of an energy function E_shape(Θ

,Σ

)=−

log({circumflex over (f)}_H) to obtain the probability of the current shape in the statistical shape model; and

computing an image term of an energy function E_image to evaluate the fit of the shape in the image

minimizing the energy function E_shape+E_image.

4. The method recited in claim 3 wherein the minimizing E_shape(Θ

,Σ

)=−

log({circumflex over (f)}_H) comprises;

computing E_shape(Θ

,Σ

_Θ

)=−

log({circumflex over (f)}_H) and E_image;

changing Θ

according to the derivative of E_shape+E_image;

generating a new covariance matrix as Σ

_Θ

, to thereby compute a new E<sub>shape(Θ

,Σ

Θ</sub>

)=−

log({circumflex over (f)}_H) and E_image until E_shape(Θ

,Σ

)+E_image is minimized.

5. The method recited in claim 1 wherein the cost function is the distance between such one of the training shapes and the transformed reference model.

6. A method for learning deformations of an object of interest obtained from an image bank of a patient having such object, comprising:

using medical imaging apparatus to perform the steps of;

generating an initial reference shape of the object of interest to be segmented;

obtaining a predetermined number of, N, images of the desired object from over a general population of such objects, where N is greater than 1;

transforming the generated reference shape to match every one of N training shapes according to an energy function comprising distorting each one of the N training shapes to overlay the reference shape with a parameter Θ

_i being a measure of the amount of distortion required at each grid point to effect the overlay for the i^th one of the N training shapes;

obtaining a vector of the parameters Θ

_i, for every one of the N training shapes through the minimization of a cost function;

estimating an uncertainty for every one of the obtained N vectors of parameters Θ

_i, such uncertainty being quantified as a covariance matrix Σ

_i;

providing a statistical model represented as {circumflex over (f)}_H(Θ

,Σ

) cumulating information of the N training shapes modeled with kernels having a mean Θ

_i, and covariance Σ

_i.

7. The method recited in claim 6 wherein the uncertainty is computed by assuming that the energy function can be approximated at its minimum by a paraboloid, the width of the paraboloid being representative of the degree of uncertainty.

8. The method recited in claim 7 wherein the reference shape is on a reference coordinate system having an overlaid free form deformation grid. Such free form deformation grid having a plurality of grid points with positions represented in the vector Θ

, with a the projection of the covariance matrix onto a grid point being represented as an ellipse elongated in the direction of greater uncertainty, and wherein, in the case of smooth linear contours, a component is tangent to the reference shape and another component is normal to the reference shape.

9. The method recited in claim 6 including additionally:

obtaining the image of the desired object from the patient;

identifying the desired object of interest in the image of the patient by positioning of the reference shape on the image comprising;

transforming the generated reference shape to match the obtained image according to the energy function comprising distorting the reference shape to overlay the obtained image with a parameter Θ

being a measure of the amount of distortion required to effect the overlay;

estimating an uncertainty, such uncertainty being quantified as a covariance matrix Σ

;

computing an energy function E_shape to obtain the probability of the current shape in the statistical shape model; and

minimizing E=E_shape+E_image comprising;

computing E_shape(Θ

,Σ

)=−

log({circumflex over (f)}_H);

computing E_image;

changing Θ

;

generating an updated covariant matrix with the changed Θ

;

recomputing the energy function E=E_shape+E_image.