Method of database-guided segmentation of anatomical structures having complex appearances

• US 7,876,934 B2
• Filed: 11/02/2005
• Issued: 01/25/2011
• Est. Priority Date: 11/08/2004
• Status: Active Grant

• Current Assignee: Siemens Healthineers AG (Siemens AG)
• Original Assignee: Siemens Medical Solutions USA Incorporated (Siemens AG)
• Inventors: Gupta, Alok, Zhou, Xiang, Georgescu, Bogdan, Rao, R. Bharat, Comaniciu, Dorin
• Primary Examiner(s): Bali; Vikkram
• Assistant Examiner(s): Brooks; Julian D
• Application Number:
  US11/265,772
• Publication Number:
  US 20060171586A1
• Time in Patent Office: 1,910 Days
• Field of Search: 382/128, 382/131, 382/132, 382/164, 382/165, 382/171, 382/173, 382/209, 382/218, 382/220, 382/224, 382/130, 382/181, 382/195, 382/199, 382/203, 382/225, 128/922, 128/923
• US Class Current: 382/128
• CPC Class Codes: G06F 18/28 Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
• CPC Class Codes: G06T 2207/10132 Ultrasound image
• CPC Class Codes: G06T 2207/20081 Training; Learning
• CPC Class Codes: G06T 2207/30048 Heart; Cardiac
• CPC Class Codes: G06T 7/11 Region-based segmentation
• CPC Class Codes: G06V 10/772 Determining representative reference patterns, e.g. averaging or distorting patterns; Generating dictionaries
• CPC Class Codes: Y10S 128/922 including image analysis
• CPC Class Codes: Y10S 128/923 by comparison of patient data to other data

Abstract

A method for segmenting an anatomical structure of interest within an image is disclosed. The anatomical structure of interest is compared to a database of images of like anatomical structures. Those database images of like anatomical structures that are similar to the anatomical structure of interest are identified. The identified database images are used to detect the anatomical structure of interest in the image. The identified database images are also used to determine the shape of the anatomical structure of interest. The anatomical structure of interest is segmented from the image.

Claims

1. A method for segmenting an organ comprising the steps of:

detecting an organ with a trained detector, wherein the step of detecting an organ comprises;

receiving a first image that includes an organ; and

applying a detector trained on a database of images of like organs to the first image to discriminate the organ in the first image from a background of the age; and

determining a most likely shape of the organ, wherein the determining step comprises;

identifying images from the database of images of like organs that are closest to the organ in the first image by using a distance between learned image features that best characterize the shape of the organ,

wherein the database of images of like organs is used to learn the image features that best characterize the shape of the organ by emulating a distance between feature vectors to a distance between shapes;

inferring the shape of the organ in the first image by using the identified database images to first determine the whole shape of the organ and then determine smaller parts of the shape; and

segmenting the organ from the first image by using the inferred shape of the organ,

wherein the method is performed using a processor, and

wherein d(c_q,c_r)=(c_q−

c_r)^T(c_q−

c_r) is the distance between shapes, d(f_q,f_r)=(f_q−

f_r)^TΣ

(f_q−

f_r) is the distance between feature vectors, where f is an appearance feature vector, c is its corresponding shape, (f_q,c_q),(f_r,c_r) represent the vector of a query and a reference, respectively, and Σ

is a linear metric associated with the feature vector space.

2. The method of claim 1 wherein the images of like organs are expert annotated organs, and

wherein the step of applying a detector comprises the steps of:

using the database to train the detector to discriminate between the organ in the first image and the background of the first image; and

using the detector to determine a location of the organ within the first image.

3. The method of claim 1 wherein the step of identifying images from the database of images of like organs comprises the steps of:

maintaining sample based representations for the database of images;

clustering the representations based on the distance between shapes;

obtaining feature scores for the organ in the first image, the feature scores being indicative of the appearance of the organ in the first image; and

selecting the features that best separate the clusters.

4. The method of claim 3 wherein features are selected that minimize the distance between the inferred shape and the actual shape of the organ in the first image.

5. The method of claim 3 wherein the features include non-image features including demographic, clinical, or genetic/genomic/proteomic information.

6. The method of claim 1 wherein the step of using the identified features to infer the shape of the organ in the first image comprises the steps of:

applying a nearest neighbor approach to identify a closest match between a set of feature scores and a set of sample representations of features and shapes of organs;

using a distance measurement to quantify similarities between the feature score set and the sample representations set; and

selecting the representative samples that are most similar to the organ in the first mage based on the distance measurement.

7. The method of claim 6 wherein a small distance measurement between features is indicative of a sample shape that matches the shape of the organ in the first image.

8. The method of claim 1 wherein the organ in the first image comprises a left ventricle.

9. The method of claim 1 wherein the first image is a two dimensional image.

10. The method of claim 1 wherein the first image is a three dimensional image.

11. The method of claim 1 wherein the images in the databases are annotated by demographic, clinical, or genetic/genomic/proteomic information.

12. The method of claim 1 wherein the closeness between images are measures with consideration of non-image features including demographic, clinical, or genetic/genomic/proteomic information.

13. The method of claim 1, wherein the detector is trained by:

using a weighted alignment scheme to increase influence of stable landmark points associated with the organ in the first image;

eliminating influences of known invalid regions in the first image by using an occlusion mask; and

using a boosted cascade of simple classifiers to train the detector on valid regions of the first image.

14. The method of claim 13 wherein the invalid regions are occlusions.

15. The method of claim 13 wherein the step of eliminating influences further comprises the steps of:

setting an intensity value for each pixel in an invalid region to zero; and

determining a number of valid pixels by computing an integral mask of the first image.

16. The method of claim 13 wherein the step of using a weighted alignment scheme comprises the steps of:

assigning higher weights to more stable points in the first image; and

using a least-square alignment scheme to normalize the shape of the first image of the organ.

17. A method for segmenting an organ comprising the steps of:

detecting an organ with a trained detector, wherein the step of detecting an organ comprises;

receiving a first image that includes an organ; and

applying a detector trained on a database of images of like organs to the first image to discriminate the organ in the first image from a background of the first image; and

determining a most likely shape of the organ, wherein the detector is trained by;

using a weighted alignment scheme to increase influence of stable landmark points associated with the organ in the first image;

eliminating influences of known invalid regions in the first image by using an occlusion mask; and

using a boosted cascade of simple classifiers to train the detector on valid regions of the first image,

wherein with a weight matrix W, a minimized criterion for aligning shapes is given by ℑ

_GPA=∥

s_iR_ic_i+t_i−

c∥

w, where c_i represents the i^th shape control point, s_i,R_i,t_i represent scale, rotation and translation, respectively, and c is the mean shape, and

wherein the method is performed using a processor.

18. A method for segmenting an organ comprising the steps of:

detecting an organ with a trained detector, wherein the step of detecting an organ comprises;

receiving a first image that includes an organ; and

applying a detector trained on a database of images of like organs to the first image to discriminate the organ in the first image from a background of the first image; and

determining a most likely shape of the organ, wherein the detector is trained by;

using a weighted alignment scheme to increase influence of stable landmark points associated with the organ in the first image;

eliminating influences of known invalid regions in the first image by using an occlusion mask; and

using a boosted cascade of simple classifiers to train the detector on valid regions of the first image,

wherein eliminating influences of known invalid regions in the first image by using an occlusion mask comprises;

determining, an intensity value I for each pixel (x, y) in the first image and, for each pixel at a location (x₀, y₀) in the first image, computing its intensity value as follows;