Online learning enhanced atlas-based auto-segmentation

Patent metadata

Field	Value
Publication number	US-10410348-B2
Application number	US-201615386673-A
Country	US
Kind code	B2
Filing date	Dec 21, 2016
Priority date	Dec 21, 2016
Publication date	Sep 10, 2019
Grant date	Sep 10, 2019

Abstract

An image segmentation method is disclosed. The method includes receiving a plurality of atlases and a subject image, each atlas including an atlas image showing a structure of interest and associated structure delineations, the subject image being acquired by an image acquisition device and showing the structure of interest. The method further includes calculating, by an image processor, mapped atlases by registering the respective atlases to the subject image, and determining, by the image processor, a first structure label map for the subject image based on the mapped atlases. The method also includes training, by the image processor, a structure classifier using a subset of the mapped atlases, and determining, by the image processor, a second structure label map for the subject image by applying the trained structure classifier to one or more subject image points in the subject image. The method additional includes combining, by the image processor, the first label map and the second label map to generate a third label map representative of the structure of interest.

First claim

What is claimed is:

1. An image segmentation method, comprising:
receiving a plurality of atlases and a subject image, each atlas including an atlas image showing a structure of interest and associated structure delineations, the subject image being acquired by an image acquisition device and showing the structure of interest;
calculating, by an image processor, mapped atlases by registering the respective atlases to the subject image;
determining, by the image processor, a first structure label map for the subject image based on the mapped atlases;
determining that a first structure label of a first set of the mapped atlases for a given region of the subject image is different from a second structure label of a different second set of the mapped atlases for the given region;
in response to determining that the first structure label is different from the second structure label, selecting training samples corresponding to the given region for a subset of the mapped atlases comprising at least one of the plurality of atlases registered to the subject image;
training, by the image processor, a structure classifier using the selected training samples corresponding to the given region;
determining, by the image processor, a second structure label map for the subject image by applying the trained structure classifier to one or more subject image points in the subject image; and
combining, by the image processor, the first label map and the second label map to generate a third label map representative of the structure of interest.

2. The method of claim 1 , wherein calculating the mapped atlases includes:
mapping the atlas image in each atlas to the subject image;
calculating a registration transformation for each atlas based on the mapping; and
calculating mapped structure delineations for each atlas by applying the registration transformation to the structure delineations of the atlas.

3. The method of claim 2 , wherein determining the first structure label map includes:
determining atlas-based auto-segmentation (ABAS) structure label maps corresponding to the atlases based on the respective mapped structure delineations; and
determining the first structure label map by fusing the ABAS structure label maps.

4. The method of claim 3 , wherein fusing the ABAS structure label is according to at least one of a majority voting method or a simultaneous truth and performance level estimation (STAPLE) method.

5. The method of claim 1 , wherein registering the respective atlases to the subject image includes mapping each atlas image and the subject image to a common reference image.

6. The method of claim 5 , wherein the reference image is an average atlas image obtained by averaging the atlas images.

7. The method of claim 1 , further comprising:
selecting the subset of the mapped atlases based on a selection criterion.

8. The method of claim 7 , wherein selecting the subset of the mapped atlases includes:
determining an image similarity between each :napped atlas image and the subject image;
ranking the mapped atlases based on the image similarities of the respective mapped atlas images; and
selecting the subset of the mapped atlases based on the ranking.

9. The method of claim 8 , wherein determining the image similarity includes determining a global similarity indicating how the corresponding mapped atlas image as a whole correlates with the subject image as a whole, or determining a local similarity representing how the structure of interest in the corresponding mapped atlas image correlates with the structure of interest in the subject image.

10. The method of claim 1 , wherein the machine learning algorithm to the one or more mapped atlases includes applying a random forest algorithm to the one or more mapped atlases to obtain a random forest model.

11. The method of claim 1 , further comprising:
selecting a plurality of training samples from each of the one or more mapped atlases; and
training the structure classifier using the plurality of training samples.

12. The method of claim 11 , wherein determining that the first structure label of the first set of the mapped atlases is different from the second structure label of the second set of the mapped atlases comprises determining that a number of mapped atlases in the first set exceeds a specified amount such that a disagreement between structure labels among the mapped atlases for the given region is larger than a certain level.

13. The method of claim 11 , further comprising:
computing attributes for the training samples,
wherein training the structure classifier uses the attributes.

14. The method of claim 13 , wherein the attributes are computed using a pre-trained convolutional neural network.

15. The method of claim 1 , further comprising:
selecting the one or more subject image points from the subject image based on the first structure label map.

16. The method of claim 15 , wherein the one or more selected subject image points correspond to structural labels indicative of segmentation ambiguity.

17. An image segmentation apparatus, comprising:
a memory configured to receive and store a plurality of atlases and a subject image, each atlas including an atlas image showing a structure of interest and associated structure delineations, the subject image being acquired by an image acquisition device and showing the structure of interest; and
an image processor coupled to the memory and configured to:
calculate mapped atlases by registering the respective atlases to the subject image;
determine a first structure label map for the subject image based on the mapped atlases;
determine that a first structure label of a first set of the mapped atlases for a given region of the subject image is different from a second structure label of a different second set of the mapped atlases for the given region;
in response to determining that the first structure label is different from the second structure label, select training samples corresponding to the given region for a subset of the mapped atlases comprising at least one of the plurality of atlases registered to the subject image;
train a structure classifier using the selected training samples corresponding to the given region;
determine a second structure label map for the subject image by applying the trained structure classifier to one or more subject image points in the subject image; and
combine the first label map and the second label map to generate a third label map representative of the structure of interest.

18. The image segmentation apparatus of claim 17 , wherein the image processor is further configured to:
select the subset of the mapped atlases based on image similarities between the respective mapped atlas images and the subject image.

19. The image segmentation apparatus of claim 17 , wherein the image processor is further configured to:
select the one or more subject image points from the subject image based on the first structure label map.

20. A non-transitory computer-readable storage medium storing instructions that, when executed by an image processor, cause the processor to perform an image segmentation method, comprising:
receiving a plurality of atlases and a subject image, each atlas including an atlas image showing a structure of interest and associated structure delineations, the subject image being acquired by an image acquisition device and showing the structure of interest;
calculating mapped atlases by registering the respective atlases to the

Assignees

• Elekta Inc

Inventors

• Han Xiao

Classifications

• G06T2207/20081

  Training; Learning · CPC title

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G06T2207/20128

  Atlas-based segmentation · CPC title

• G06T7/11 Primary

  Region-based segmentation · CPC title

• G06T7/0012

  Biomedical image inspection · CPC title