Region clustering forest for analyzing medical imaging data

Patent metadata

Field	Value
Publication number	US-9576356-B2
Application number	US-201514707418-A
Country	US
Kind code	B2
Filing date	May 8, 2015
Priority date	May 8, 2015
Publication date	Feb 21, 2017
Grant date	Feb 21, 2017

Abstract

Systems and methods for training a region clustering forest include receiving a set of medical training images for a population of patients. A set of image patches is extracted from each image in the set of medical training images. A plurality of region clustering trees are generated each minimizing a loss function based on respective randomly selected subsets of the set of image patches to train the region clustering forest. Each of the plurality of region clustering trees cluster image patches at a plurality of leaf nodes and the loss function measures a compactness of the cluster of image patches at each leaf node in each of the plurality of region clustering trees.

First claim

The invention claimed is:

1. A method for training a region clustering forest, comprising:
receiving a set of medical training images for a population of patients;
extracting a set of image patches from each image in the set of medical training images; and
generating a plurality of region clustering trees each minimizing a loss function based on respective randomly selected subsets of the set of image patches to train the region clustering forest, wherein each of the plurality of region clustering trees cluster image patches at a plurality of leaf nodes and the loss function measures a compactness of the cluster of image patches at each leaf node in each of the plurality of region clustering trees.

2. The method as recited in claim 1 , wherein extracting a set of image patches from each image in the set of medical training images comprises:
extracting image patches centered around every n voxels in each image of the set of medical training images, where n is any positive integer.

3. The method as recited in claim 1 , wherein generating a plurality of region clustering trees each minimizing a loss function based on respective randomly selected subsets of the set of image patches to train the region clustering forest comprises:
for each region clustering tree of the plurality of region clustering trees:
randomly selecting a respective subset of the set of image patches at a root node of the respective region clustering tree;
selecting a classifier for a current node of the respective region clustering tree resulting in a classification of image patches at the current node that minimizes the loss function for the respective region clustering tree;
sorting image patches at the current node using the selected classifier into a plurality of child nodes of the current node; and
repeating the selecting the classifier and the sorting for all resulting child nodes until a stopping criterion is satisfied.

4. The method as recited in claim 3 , wherein repeating the selecting the classifier and the sorting for all resulting child nodes until a stopping criterion is satisfied comprises:
repeating the selecting the classifier and the sorting until a predetermined maximum depth of the respective region clustering tree is reached.

5. The method as recited in claim 1 , wherein generating a plurality of region clustering trees each minimizing a loss function based on respective randomly selected subsets of the set of image patches to train the region clustering forest comprises:
generating each of the plurality of region clustering trees to generate different minimizations of the loss function based on the respective randomly selected subsets.

6. The method as recited in claim 1 , further comprising:
determining offset vectors between a center of each image patch in the set of image patches and a known location of a landmark in the set of medical training images; and
computing an associated offset vector for each respective leaf node of each of the plurality of region clustering trees based on the determined offset vectors for image patches sorted to the respective leaf node.

7. The method as recited in claim 6 , wherein computing an associated offset vector for each respective leaf node of each of the plurality of region clustering trees based on the determined offset vectors for image patches sorted to the respective leaf node comprises at least one of:
computing the associated offset vector for each respective leaf node as a mean of the determined offset vectors for image patches sorted to the respective leaf node;
computing the associated offset vector for each respective leaf node as a median of the determined offset vectors for image patches sorted to the respective leaf node; and
computing the associated offset vector for each respective leaf node by machine-learning based clustering of the determined offset vectors for image patches sorted to the respective leaf node.

8. The method as recited in claim 6 , further comprising:
extracting patches from medical imaging data including a landmark of a patient;
sorting the extracted patches to a respective one of the leaf nodes for each of the plurality of region clustering trees of the trained region clustering forest; and
determining a location of the landmark based on the associated offset vectors associated with the respective one of the leaf nodes that the extracted patches were sorted to.

9. The method as recited in claim 1 , wherein generating a plurality of region clustering trees each minimizing a loss function based on respective randomly selected subsets of the set of image patches to train the region clustering forest comprises:
associating a label with each leaf node of each of the plurality of region clustering trees indicating a presence of bone based on annotations in the medical training images.

10. The method as recited in claim 9 , further comprising:
extracting patches from medical imaging data including bone of a patient;
sorting the extracted patches to a respective one of the leaf nodes for each of the plurality of region clustering trees of the trained region clustering forest; and
determining whether the extracted patches include the bone based on the label associated with the respective one of the leaf nodes that the extracted patches were sorted to.

11. The method as recited in claim 1 , further comprising:
annotating an atlas of a target anatomical object with atlas labels;
extracting atlas patches from the atlas;
sorting the extracted atlas patches to a respective one of the leaf nodes for each of the plurality of region clustering trees of the trained region clustering forest; and
associating an index of the respective one of the leaf nodes to voxel positions in the atlas associated with extracted atlas patches sorted to the respective one of the leaf nodes.

12. The method as recited in claim 11 , further comprising:
extracting patches from medical imaging data including the target anatomical object of a patient;
sorting the extracted patches to another respective one of the leaf nodes for each of the plurality of region clustering trees of the trained region clustering forest; and
assigning labels to voxels associated with the extracted patches based on the atlas labels associated with the index of the leaf nodes that the extracted patches were sorted to.

13. An apparatus for training a region clustering forest, comprising:
means for receiving a set of medical training images for a population of patients;
means for extracting a set of image patches from each image in the set of medical training images; and
means for generating a plurality of region clustering trees each minimizing a loss function based on respective randomly selected subsets of the set of image patches to train the region clustering forest, wherein each of the plurality of region clustering trees cluster image patches at a plurality of leaf nodes and the loss function measures a compactness of the cluster of image patches at each leaf node in each of the plurality of region clustering trees.

14. The apparatus as recited in claim 13 , wherein the means for extracting a set of image patches from each image in the set of medical training images comprises:
means for extracting image patches centered around every n voxels in each image of the set of medical training images, where n is any positive integer.

15. The apparatus as recited in claim 13 , wherein the means for generating a plurality of region clustering trees each minimizing a loss function based on respective randomly selected subsets of the set of image patches to train the region

Assignees

• Siemens Healthcare Gmbh

Inventors

• Lay Nathan
• Yang Dong
• Liu David
• Zhou Shaohua Kevin

Classifications

• G06V10/774

  Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06V10/772

  Determining representative reference patterns, e.g. averaging or distorting patterns; Generating dictionaries · CPC title

• G06V10/762

  using clustering, e.g. of similar faces in social networks · CPC title

• G06T7/337 Primary

  involving reference images or patches · CPC title

• G06T7/0012 Primary

  Biomedical image inspection · CPC title