System and method for improved medical images

Patent metadata

Field	Value
Publication number	US-11164308-B2
Application number	US-201815906005-A
Country	US
Kind code	B2
Filing date	Feb 27, 2018
Priority date	Feb 27, 2017
Publication date	Nov 2, 2021
Grant date	Nov 2, 2021

Abstract

A method for using machine learning to perform classification of anatomical coverage of images includes acquiring a series of medical images of a subject. The method also includes automatically, with a computer system, analyzing each image in the series of medical images using a machine-learning technique to classify each image in the series of medical images based on anatomical structures reflected in each image in the series of medical images.

First claim

The invention claimed is:

1. A system comprising:
at least one medical imaging device acquiring at least one medical image, comprising at least a training scan and an input scan, from a subject;
a computing device coupled to a network and comprising at least one processor executing computer executable instructions that, when executed, cause the system to:
receive the training scan;
perform a feature extraction from a series of acquired medical images, comprising:
selecting a template aligned with the training scan;
splitting the training scan into a set of kxkxk even and non-overlapping blocks; and
generating a training data for the training scan comprising a feature vector derived from the set of kxkxk even and non-overlapping blocks;

receive the input scan;
select a subset of features for:
the input scan; and
a plurality of training scans including the training scan;

classify the input scan as a brain image, a chest image, an abdomen-pelvis image, or a chest-abdomen-pelvis image according to at least one machine learning technique comparing the features for the input scan with the feature vector for each of the plurality of training scans, wherein a classification of the input scan is classified automatically, according to an image-covering classification comprising an overall coverage of the at least one medical image or an associated image view, based on an anatomy.

2. The system of claim 1 , wherein the at least one medical imaging device comprises a computed tomography (CT) or magnetic resonance imaging (MRI) device.

3. The system of claim 1 , wherein:
the template comprises a scan from a positively classified sample selected from a training set; and
aligning the template with the training scan comprises a multiresolution affine registration involving three levels, with a mutual information as a cost function.

4. The system of claim 1 , wherein the mean intensity in each of the plurality of non-overlapping blocks is computed to represent a corresponding block.

5. The system of claim 4 , wherein the subset of features is selected using a correlation-based feature selection (CFS) algorithm comprising a filter based feature selection method.

6. The system of claim 1 , wherein the subset of features is selected based on a heuristic merit, taking into account:
at least one individual feature for predicting a class label; and
a level of inter-correlation among the subset of features.

7. The system of claim 1 , wherein the input-scan_is classified utilizing:
a label generated in association with the training scan in the plurality of training scans; and
a feature vector within the training data for each of the plurality of training scans.

8. The system of claim 1 , wherein the machine learning technique:
is trained by comparing a first plurality of features in the plurality of training scans with a second plurality of features in a plurality of binary clusters comprising a plurality of positively classified samples and a plurality of negatively classified samples;
employs a one-vs-rest strategy;
classifies, without user interaction, the at least one medical image according to an overall coverage of the anatomical structures reflected in the at least one medical image.

9. T he system of claim 1 , wherein the machine learning technique is a support vector machine (SVM) construct based on a radial basis function (RBF) kernel used to build a classification model.

10. A method comprising:
receiving, by a computing device coupled to a network and comprising at least one processor executing computer executable instruction within a memory, a training scan from at least one medical imaging device acquiring at least one medical image from a subject;
performing, by the computing device, a feature extraction from a series of acquired medical images, comprising:
selecting a template aligned with the training scan;
splitting the training scan into a set of kxkxk even and non-overlapping blocks; and
generating a training data for the training scan comprising a feature vector derived from the set of kxkxk even and non-overlapping blocks;

receiving, by the computing device, an input scan;
selecting, by the computing device, a subset of features for:
the input scan; and
a plurality of training scans including the training scan;

classifying, by the computing device, the input scan as a brain image, a chest image, an abdomen-pelvis image, or a chest-abdomen-pelvis image according to at least one machine learning technique comparing the features for the input scan with the feature vector for each of the plurality of training scans, wherein a classification of the input scan is classified automatically, according to an image-covering classification comprising an overall coverage of the at least one medical image or an associated image view, based on an anatomy.

11. The method of claim 10 , wherein the at least one medical imaging device comprises a computed tomography (CT) or magnetic resonance imaging (MRI) device.

12. The method of claim 10 , wherein:
the template comprises a scan from a positively classified sample selected from a training set; and
aligning the template with the training scan comprises a multiresolution affine registration involving three levels, with a mutual information as a cost function.

13. The method of claim 10 , wherein the mean intensity in each of the plurality of non-overlapping blocks is computed to represent a corresponding block.

14. The method of claim 13 , further comprising the step of selecting the subset of features using a correlation-based feature selection (CFS) algorithm comprising a filter based feature selection method.

15. The method of claim 10 , further comprising the step of selecting the subset of features based on a heuristic merit, taking into account:
at least one individual feature for predicting a class label; and
a level of inter-correlation among the subset of features.

16. The method of claim 10 , wherein the input-scan is classified utilizing:
a label generated in association with the training scan in the plurality of training scans; and
a feature vector within the training data for each of the plurality of training scans.

17. The method of claim 10 , wherein the machine learning technique:
is trained by comparing a first plurality of features in the plurality of training scans with a second plurality of features in a plurality of binary clusters comprising a plurality of positively classified samples and a plurality of negatively classified samples;
employs a one-vs-rest strategy;
classifies, without user interaction, the at least one medical image according to an overall coverage of the anatomical structures reflected in the at least one medical image.

18. The method of claim 10 , wherein the machine learning technique is a support vector machine (SVM) construct based on a radial basis function (RBF) kernel used to build a classification model.

Assignees

• Univ California

Inventors

• Wang Xiaoyong
• Lo Pechin
• Brown Matthew
• Ramakrishna Bharath
• Goldin Jonathan

Classifications

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06F18/2411

  based on the proximity to a decision surface, e.g. support vector machines · CPC title

• G06F18/2433

  Single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection · CPC title

• G06F18/214

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

• G06V10/751

  Comparing pixel values or logical combinations thereof, or feature values having positional relevance, e.g. template matching · CPC title