System and method for lesion detection using locally adjustable priors
View Patent ↗According to an aspect of the invention, a method for training a classifier for classifying candidate regions in computer aided diagnosis of digital medical images includes providing a training set of annotated images, each image including one or more candidate regions that have been identified as suspicious, deriving a set of descriptive feature vectors, where each candidate region is associated with a feature vector. A subset of the features are conditionally dependent, and the remaining features are conditionally independent. The conditionally independent features are used to train a naïve Bayes classifier that classifies the candidate regions as lesion or non-lesion. A joint probability distribution that models the conditionally dependent features, and a prior-odds probability ratio of a candidate region being associated with a lesion are determined from the training images. A new classifier is formed from the naïve Bayes classifier, the joint probability distribution, and the prior-odds probability ratio.
1. A method for training a classifier for classifying candidate regions in computer aided diagnosis of digital medical images, said method comprising the steps of:
providing a training set of images, each said image including one or more candidate regions that have been identified as suspicious by a candidate generation step of a computer aided diagnosis system, and wherein each said image has been manually annotated to identify lesions;
deriving a set of descriptive feature vectors from a feature computation step of a computer aided diagnosis system, wherein each candidate region is associated with a feature vector, wherein a subset of said features are conditionally dependent, and the remaining features are conditionally independent;
using said conditionally independent features to train a naïve Bayes classifier that classifies said candidate regions as lesion or non-lesion;
determining a joint probability distribution from said training images that models the conditionally dependent features;
determining from said training images a prior-odds probability ratio of a candidate region being associated with a lesion; and
forming a new classifier from a product of said naïve Bayes classifier, said joint probability distribution for the conditionally dependent features, and said prior-odds probability ratio.
2. The method of claim 1 , further comprising determining an operator threshold of said classifier from an ROC curve of predictions of said classifier.
3. The method of claim 1 , wherein said joint probability distribution for the conditionally dependent features is a ratio of whether or not a candidate region is a lesion given said conditionally dependent features, and wherein determining said joint probability distribution comprises using a kernel-density estimation to estimate a distribution for said candidate region to be a lesion, and to estimate a distribution for said candidate region to be a non-lesion.
4. The method of claim 1 , wherein said conditionally dependent features incorporate information regarding a spatial location of said candidate region in an object of interest.
5. The method of claim 4 , wherein said object of interest is a colon, and said spatial location is a normalized distance from a rectum measured along a colon centerline.
6. The method of claim 4 , wherein said object of interest is a pair of lungs, and said spatial location is specified in a lung coordinate system.
7. The method of claim 6 , wherein said lung coordinate system is specified in terms of a distance from a lung center reference or hilum, a distance from a lung wall, a distance from a lung apex or basal point or carina, and where it is specified whether said spatial location is in a left lung or a right lung.
8. The method of claim 6 , wherein said lung coordinate system is specified in terms of distances from key landmarks, vessels and airway tree structures in and around the lungs.
9. A method for training a classifier for classifying candidate regions in computer aided diagnosis of digital medical images, said method comprising the steps of:
providing a training set of images, each said image including one or more candidate regions that have been identified as suspicious by a candidate generation step of a computer aided diagnosis system, and wherein each said image has been manually annotated to identify lesions;
deriving a set of descriptive feature vectors from a feature computation step of a computer aided diagnosis system, wherein each candidate region is associated with a feature vector, wherein a subset of set features are conditionally dependent and the remaining features are conditionally independent, said conditionally dependent features incorporate information regarding a spatial location and physical attributes of said candidate region in an object of interest; and
training a classifier using a Bayesian network that incorporates said conditionally dependent local spatial and physical features into a prior probability, wherein said classifier is adapted to classifying candidate regions as lesion or non-lesion.
10. The method of claim 9 , wherein said conditionally dependent physical features include one or more of a density, size, or abnormality bias of said candidate region.
11. The method of claim 10 , wherein said classifier is represented as
p
(
D
=
1
|
f
1
,
f
2
)
p
(
D
=
0
|
f
1
,
f
2
)
=
p
(
D
=
1
|
f
1
)
p
(
D
=
0
|
f
1
)
×
p
(
D
=
1
|
f
2
)
1
-
p
(
D
=
1
|
f
2
)
×
p
(
D
=
0
)
p
(
D
=
1
)
,
wherein p is a probability density, D=1 represents a case of a candidate region being lesion, D=0 represents a case of a candidate region being non-lesion, f 1 represents said conditionally dependent features, f 2 represents said conditionally independent features.
12. A program storage device readable by a computer, tangibly embodying a program of instructions executable by the computer to perform the method steps for classifying candidate regions in computer aided diagnosis of digital medical images, said method comprising the steps of:
providing a training set of images, each said image including one or more candidate regions that have been identified as suspicious by a candidate generation step of a computer aided diagnosis system, and wherein each said image has been manually annotated to identify lesions;
deriving a set of descriptive feature vectors from a feature computation step of a computer aided diagnosis system, wherein each candidate region is associated with a feature vector, wherein a subset of said features are conditionally dependent, and the remaining features are conditionally independent;
using said conditionally independent features to train a naïve Bayes classifier that classifies said candidate regions as lesion or non-lesion;
determining a joint probability distribution from said training images that models the conditionally dependent features;
determining from said training images a prior-odds probability ratio of a candidate region being associated with a lesion; and
forming a new classifier from a product of said naïve Bayes classifier, said joint probability distribution for the conditionally dependent features, and said prior-odds probability ratio.
13. The computer readable program storage device of claim 12 , the method further comprising determining an operator threshold of said classifier from an ROC curve of predictions of said classifier.
14. The computer readable program storage device of claim 12 , wherein said joint probability distribution for the conditionally dependent features is a ratio of whether or not a candidate region is a lesion given said conditionally dependent features, and wherein determining said joint probability distribution comprises using a kernel-density estimation to estimate a distribution for said candidate region to be a lesion, and to estimate a distribution for said candidate region to be a non-lesion.
15. The computer readable program storage device of claim 12 , wherein said conditionally dependent features incorporate information regarding a spatial location of said candidate region in an object of interest.
16. The computer readable program storage device of claim 15 , wherein said object of interest is a colon, and said spatial location is a normalized distance from a rectum measured along a colon centerline.
17. The computer readable program storage device of claim 15 , wherein said object of interest is a pair of lungs, and said spatial location is specified in a lung coordinate system.
18. The computer readable program storage device of claim 17 , wherein said lung coordinate system is specified in terms of a distance from a lung center reference or hilum, a distance from a lung wall, a distance from a lung apex or basal point or carina, and where it is specified whether said spatial location is in a left lung or a right lung.
19. The computer readable program storage device of claim 17 , wherein said lung coordinate system is specified in terms of distances from key landmarks, vessels and airway tree structures in and around the lungs.