Sequential learning technique for medical image segmentation

Patent metadata

Field	Value
Publication number	US-2017358075-A1
Application number	US-201615178511-A
Country	US
Kind code	A1
Filing date	Jun 9, 2016
Priority date	Jun 9, 2016
Publication date	Dec 14, 2017
Grant date	—

Abstract

Sequential learning techniques, such as auto-context, that apply the output of an intermediate classifier as contextual features for its subsequent classifier have shown impressive performance for semantic segmentation. It is shown that these methods can be interpreted as an approximation technique derived from a Bayesian formulation. To improve the effectiveness of applying this approximation technique, a new sequential learning approach is proposed for semantic segmentation that solves a segmentation problem by breaking it into a series of simplified segmentation problems. Sequentially solving each of the simplified problems along the path leads to a more effective way for solving the original segmentation problem. To achieve this goal, a learning-based method is proposed to generate simplified segmentation problems by explicitly controlling the complexities of the modeling classifiers. Promising results were reported on the 2013 SATA canine leg muscle segmentation dataset.

First claim

1 . A medical image segmentation method for locating an anatomical structure in an image of interest to a user, using a training set of one or more annotated, anatomical images, the method comprising:
(a) decomposing each annotated, anatomical image in the training set into a plurality of regions; (b) generating, for each annotated, anatomical image in the training set, a hierarchy that includes a plurality of levels, each level in the hierarchy corresponding to one of the plurality of regions; (c) obtaining, from storage, at least one previously derived annotation for each level in the hierarchy and training a plurality of first classifiers to approximate the previously derived annotations, wherein the first classifiers are trained in a particular sequence such that the output of one of the first classifiers becomes input to another one of the first classifiers; (d) storing the trained, first classifiers of (c); and (e) locating an anatomical structure in the image of interest to the user based on the stored trained, first classifiers.

2 . The medical image segmentation method of claim 1 , wherein the method comprises:
training, for each anatomical image in the training set, a plurality of second classifiers to approximate one or more annotations; and storing the annotations approximated using the second classifiers as the previously derived annotations.

3 . The medical image segmentation method of claim 1 , wherein a corrective learning technique is applied to the output of one of the first classifiers in order to correct annotation errors before becoming input to said another one of the first classifiers.

4 . The medical image segmentation method of claim 1 , wherein a leave-one-out cross-validation strategy is used to produce classification maps for each training image.

5 . The medical image segmentation method of claim 1 , wherein each anatomical image in the training set is decomposed into a plurality of regions based on a maximum a posteriori (MAP) segmentation procedure.

6 . The medical image segmentation method of claim 1 , wherein sequential training in (c) is based on an approximation technique derived from a Bayesian approximation of a probabilistic distribution of annotations.

7 . A medical image segmentation method for locating an anatomical structure in an image of interest to a user, using a training set of one or more anatomical images, the method comprising:
(a) decomposing each anatomical image in the training set into a plurality of regions; (b) generating, for each anatomical image in the training set, a hierarchy that includes a plurality of levels, each level in the hierarchy corresponding to one of the plurality of regions; (c) training, for each anatomical image in the training set, a plurality of first classifiers to approximate one or more annotations, and storing one or more such estimated annotations by the first classifiers as previously derived annotations; (d) training a plurality of second classifiers to approximate the previously derived annotations, where the second classifiers are trained in a particular sequence such that the output of one of the second classifiers becomes input to another one of the second classifiers; (e) storing the trained, second classifiers in (d); and (f) locating an anatomical structure in the image of interest to the user based on stored trained, second classifiers.

8 . The medical image segmentation method of claim 7 , wherein a corrective learning technique is applied to the output of one of the second classifiers in order to correct annotation errors before becoming input to said another one of the second classifiers.

9 . The medical image segmentation method of claim 7 , wherein a leave-one-out cross-validation strategy is used to produce classification maps for each training image.

10 . The method of claim 7 , wherein each anatomical image in the training set is decomposed into a plurality of regions based on a maximum a posteriori (MAP) segmentation procedure.

11 . The method of claim 7 , wherein sequential training in (d) is based on an approximation technique derived from a Bayesian approximation of a probabilistic distribution of annotations.

12 . An article of manufacture comprising a non-transitory computer storage medium storing computer readable program code which, when executed by a computer, implements a computer-based medical image segmentation method for locating an anatomical structure in an image of interest to a user, using a training set of one or more annotated, anatomical images, the method comprising:
(a) computer readable program code decomposing each annotated, anatomical image in the training set into a plurality of regions; (b) computer readable program code generating, for each annotated, anatomical image in the training set, a hierarchy that includes a plurality of levels, each level in the hierarchy corresponding to one of the plurality of regions; (c) computer readable program code obtaining, from storage, at least one previously derived annotation for each level in the hierarchy and training a plurality of first classifiers to approximate the previously derived annotations, wherein the first classifiers are trained in a particular sequence such that the output of one of the first classifiers becomes input to another one of the first classifiers; (d) computer readable program code storing the trained, first classifiers of (c); and (e) computer readable program code locating an anatomical structure in the image of interest to the user based on stored trained, first classifiers.

13 . The article of manufacture of claim 12 , wherein a corrective learning technique is applied to the output of one of the first classifiers in order to correct annotation errors before becoming input to said another one of the first classifiers.

14 . The article of manufacture of claim 12 , wherein the deriving step further comprises:
computer readable program code training a plurality of second classifiers to approximate one or more annotations for each anatomical image in the training set; and computer readable program code storing the annotations approximated using the second classifiers as the previously derived annotations.

15 . The article of manufacture of claim 12 , wherein a leave-one-out cross-validation strategy is used to produce classification maps for each training image.

16 . The article of manufacture of claim 12 , wherein each anatomical image in the training set is decomposed into a plurality of regions based on a maximum a posteriori (MAP) segmentation procedure.

17 . The article of manufacture of claim 12 , wherein sequential training in (c) is based on an approximation technique derived from a Bayesian approximation of a probabilistic distribution of annotations.

18 . An article of manufacture comprising non-transitory computer storage medium storing computer readable program code which, when executed by a computer, implements a computer-based medical image segmentation method for locating an anatomical structure in an image of interest to a user, using a training set of one or more anatomical images, the method comprising:
(a) computer readable program code decomposing each anatomical image in the training set into a plurality of regions; (b) computer readable program code generating, for each anatomical image in the training set, a hierarchy that includes a plurality of levels, each level in the hierarchy corresponding to one of the plurality of regions; (c) computer readable program code training, for each anatomical image in

Assignees

• IBM

Inventors

• Cao Yu
• Syeda-Mahmood Tanveer
• Wang Hongzhi

Classifications

• G06T2207/20081

  Training; Learning · CPC title

• G06T2207/30004

  Biomedical image processing · CPC title

• G06T7/143

  involving probabilistic approaches, e.g. Markov random field [MRF] modelling · CPC title

• G16H30/40

  for processing medical images, e.g. editing · CPC title

• G06V10/764

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