Image comparison tool tolerant to deformable image matching

Patent metadata

Field	Value
Publication number	US-9767380-B2
Application number	US-201514685307-A
Country	US
Kind code	B2
Filing date	Apr 13, 2015
Priority date	Apr 13, 2015
Publication date	Sep 19, 2017
Grant date	Sep 19, 2017

Abstract

An apparatus and method for determining an image similarity based on image features. In one aspect, the image similarity determination is based on an image comparison tool. The image comparison tool may be trained, by a machine-learning system, to estimate a similarity between images based on a subset of image data comprised by image features. The estimate may be an estimate of how similar structures found in the images would be following a geometric transformation of some of the structures. In one aspect, an atlas image for performing automatic segmentation of an image is determined according to a comparison made using the image comparison tool.

First claim

What is claimed is:

1. A method of training a machine learning system to generate an image comparison tool, the method comprising:
accessing a first medical image including first pixel data depicting first structures;
accessing a second medical image including second pixel data depicting second structures;
deforming the second medical image using a deformation vector, wherein deforming the second medical image causes the second pixel data to more closely match the first pixel data;
establishing a similarity value based on a comparison of the first and second pixel data;
extracting a first features group from a first subset of the first pixel data, and a second features group from a second subset of the second pixel data;
training a machine learning system by using the first and second features groups as inputs to the machine learning system and using the similarity value as an expected output of the machine learning system; and
generating an image comparison tool based on the training.

2. The method according to claim 1 , wherein the comparison of the first and second pixel data is a geometric comparison of the first and second structures, and further wherein the geometric comparison is made following a geometric transformation of at least one of the first and second structures.

3. The method according to claim 2 , wherein the geometric transformation comprises aligning the first and second medical images according to a deformation field, and wherein the deformation field is generated to provide an optimal match between the first structure and the second structure.

4. The method according to claim 2 , wherein the first subset and the second subset are selected from pixel data not subjected to geometric transformation.

5. The method according to claim 1 , wherein the first structures and the second structures are determined according to automatic segmentation.

6. The method according to claim 1 , wherein the first and second structures comprise first and second image anatomical structures, respectively.

7. The method according to claim 6 , wherein establishing the similarity value comprises comparing first and second image treatment plan geometries based on the first and second image anatomical structures, respectively.

8. The method according to claim 6 , wherein establishing the similarity value comprises comparing first and second image radiation dose distributions based on the first and second image anatomical structures, respectively.

9. The method according to claim 1 , wherein the first and second features groups comprise first and second image anatomical structures, respectively.

10. The method according to claim 1 , wherein the first and second images are comprised by a set of training images, and wherein the training of the machine learning system is performed using a plurality of unique pairings of training images in the set.

11. The method according to claim 10 , wherein a plurality of feature categories is comprised by both the first and second features groups, and wherein each feature category has a corresponding weight determined by the training of the machine learning system.

12. The method according to claim 1 , wherein at least one of the first image and the second image is obtained using X-ray radiography, computed tomography (CT) imaging, cone-beam computer tomography (CBCT) imaging, magnetic resonance imaging (MRI), positron emission tomography (PET) imaging, single photon emission computed tomography (SPECT) imaging, or ultrasound (US) imaging.

13. A method of selecting an atlas for automatically segmenting a medical image, the method comprising:
accessing a first medical image including first pixel data depicting first structures;
extracting a first features group from a first subset of the first pixel data;
accessing a set of reference features groups;
deforming the first medical image using a deformation vector, wherein deforming the first medical image causes the first pixel data to more closely match the set of reference features groups;
based upon a comparison using an image comparison tool, determining a similarity between the first features group and selected groups of the set of reference features groups;
based on the similarity, identifying a group of the set of reference features groups; and
selecting an atlas based on the identified group of the set of reference features groups.

14. The method according to claim 13 , wherein the image comparison tool comprises a set of weighted feature categories, wherein the weight associated with each category of the set represents a relevance for that category, and wherein the comparison is made among corresponding categories.

15. The method according to claim 13 , wherein each group of the set of reference features groups is determined from a corresponding reference image, and wherein the atlas is selected according to the corresponding reference image for the identified group of reference image features.

16. The method according to claim 13 , wherein a plurality of atlases is selected based on an identified plurality of groups selected from the set of reference features groups.

17. The method according to claim 13 , wherein the image comparison tool is generated by a machine learning system, wherein the machine learning system is trained by:
accessing a first training image including first training pixel data depicting first training structures;
accessing a second training image including second training pixel data depicting second training structures;
establishing the similarity value based on a comparison of the first and second training pixel data;
extracting a first training features group from a first subset of the first training pixel data, and a second training features group from a second subset of the second training pixel data; and
training the machine learning system by using the first and second training features groups as inputs to the machine learning system and using the similarity value as an expected output of the machine learning system.

18. An apparatus for determining a similarity between tomogram images, the apparatus comprising:
a computing device comprising a processor that accesses data representing a first tomogram image including first pixel data depicting a first features group, accesses data representing a set of reference features groups corresponding to a set of reference images, deforms the first tomogram image using a deformation vector, wherein deforming the first tomogram image causes the first pixel data to more closely match the set of reference features groups, determines a similarity between the first features group and a selection of groups of the set of reference features groups, and based on the similarity, and selects at least one group of the set of reference features groups to determine an atlas for automatic segmentation,
wherein the similarity determination is based on a comparison of the features groups using an image comparison tool.

19. The apparatus according to claim 18 , wherein a plurality of atlases is selected based on an identified plurality of groups selected from the set of reference features groups.

Assignees

• Varian Medical Systems Int Ag

Inventors

• Haas Benjamin
• Coradi Thomas
• Waschbuesch Michael

Classifications

• G06V10/776 Primary

  Validation; Performance evaluation · CPC title

• G06F18/217

  Validation; Performance evaluation; Active pattern learning techniques · CPC title

• G06T2207/30004

  Biomedical image processing · CPC title

• G06T2207/20128

  Atlas-based segmentation · CPC title

• G06T7/149

  involving deformable models, e.g. active contour models · CPC title