Method and system of providing a probability distribution to aid the detection of tumors in mammogram images

• US 8,194,965 B2
• Filed: 11/19/2007
• Issued: 06/05/2012
• Est. Priority Date: 11/19/2007
• Status: Active Grant

• Current Assignee: Parascript, LLC
• Original Assignee: Parascript, LLC
• Inventors: Lossev, Ilia, Bogotskaya, Natasha
• Primary Examiner(s): Koziol, Stephen R
• Application Number:
  US11/942,641
• Publication Number:
  US 20090129655A1
• Time in Patent Office: 1,660 Days
• Field of Search: 382/100, 382/132, 382/224
• US Class Current: 382/132
• CPC Class Codes: G06F 18/295 Markov models or related models, e.g. semi-Markov models; Markov random fields; Networks embedding Markov models
• CPC Class Codes: G06T 2207/10116 X-ray image
• CPC Class Codes: G06T 2207/20021 Dividing image into blocks, subimages or windows
• CPC Class Codes: G06T 2207/20076 Probabilistic image processing
• CPC Class Codes: G06T 2207/30068 Mammography; Breast
• CPC Class Codes: G06T 7/0012 Biomedical image inspection

Abstract

Methods and systems are disclosed to aid in the detection of cancer or lesion in a mammogram images. Two mammogram images are input into an application that aids in determining the probability of a cancer or lesion being present in one or both of the images. The images are divided into different nodes and labels are applied to the nodes. The first node is compared to different variants of corresponding nodes on the second image as well as neighboring nodes on the first image. Based upon the comparisons, a unary and binary potential is calculated for the label that is applied to the node. The process is repeated for every possible label and for every node. Once the unary and binary potentials have been calculated, the potentials are input into a Conditional Random Field model to determine the probability of cancer for each node of the images.

Claims

1. A computer implemented method for determining a probability distribution of a classification of labels for nodes of a first image by comparing the first image to a second image, the computer implemented method comprising:

a. defining a set of labels for each node in a plurality of nodes on the first image wherein a label of the set of labels indicates a corresponding node on the second image, a classification of a current node on the first image, and a classification of the corresponding node on the second image;

b. calculating local descriptive information for multiple mappings between, and multiple classification assignments of, the plurality of nodes of the first image and corresponding nodes on the second image;

c. inputting at least some of the descriptive information into a probabilistic model, wherein the probabilistic model accounts for classification assignments of the current node combined with different variants of corresponding nodes on the second image;

d. determining one or more probabilities of classification for at least some of the nodes; and

e. providing the determined probabilities of classification.

2. The computer implemented method of claim 1, wherein calculating descriptive information further comprises determining a unary potential for at least some of the plurality of nodes of the first image, wherein the unary potential is dependent upon labels applied to the plurality of nodes and determining a binary potential between at least some of the plurality of nodes, wherein the binary potential is dependent upon a pair of labels.

3. The computer implemented method of claim 2, wherein determining a unary potential for labels applied to at least some of the plurality of nodes further comprises:

determining the unary potential of a node on the first image if the node or the corresponding node on the second image is labeled as cancer; and

determining the unary potential of the node on the first image if the node and the corresponding node on the second image are labeled as normal.

4. The computer implemented method of claim 3, wherein determining the unary potential of the node on the first image if the node or the corresponding node on the second image is labeled as cancer further comprises:

comparing values of features associated with the node and the corresponding node on the second image with typical cancer features; and

penalizing a label assigned to the node if the feature values are atypical.

5. The computer implemented method of claim 4, wherein the features associated with a node comprise at least one of an intensity, a gradient of the intensity, and a contrast.

6. The computer implemented method of claim 3, wherein determining the unary potential of the node on the first image if the node and the corresponding node on the second image are labeled as normal further comprises:

comparing features associated with the node on the first image and the corresponding node on the second image; and

penalizing the label applied to the first node for dissimilarity between the features associated with the first node and the second node.

7. The computer implemented method of claim 6, wherein the features associated with a node comprise at least one of an intensity, a gradient of the intensity, and a contrast.

8. The computer implemented method of claim 2, wherein determining the probability distribution further comprises computing the binary potential for a pair of neighboring nodes in the first image and a pair of labels applied to the pair of neighboring nodes.

9. The computer implemented method of claim 8, wherein computing the binary potential further comprises:

penalizing the pair of labels applied to the neighboring nodes for dissimilarity in the mapping parts of the pair labels; and

penalizing the pair of labels applied to the pair of neighboring nodes for dissimilarity in the classification parts of the pair of labels.

10. The computer implemented method of claim 1, wherein the probabilistic model is a Conditional Random Field probabilistic model.

11. The computer implemented method of claim 1, wherein the step of determining one or more probabilities comprises estimating Posterior Marginal Probability for each label applied to each node of the first image in the framework of a Conditional Random Field model.

12. The computer implemented method of claim 1, wherein the step of determining probabilities comprises calculating a Posterior Marginal Probability of a classification in each node by marginalizing over all possible mappings and over all possible classifications of corresponding nodes on the other image.

13. A computer storage media not consisting of a propagated data signal, the computer storage media encoding a computer program of instructions that, when executed by at least one processor, perform a method of determining a probability distribution of labels for first and second mammogram images, the method comprising:

determining a probability distribution of labels for nodes in the first image, wherein the step of determining the probability distribution comprises;

a. defining a set of possible labels of a plurality of nodes on the first image, wherein each label comprises a classification part and a mapping part;

b. determining a unary potential for each node of the first image, wherein determining the unary potential comprises;

i. determining the unary potential of a first node on the first image for labels classifying the first node or a corresponding node on the second image as cancer, wherein the unary potential is derived by comparing values of features associated with the first node on the first image or the corresponding node on the second image with at least one known value and penalizing a label assigned to the first node if feature values are atypical compared to the at least one known value; and

ii. determining the unary potential of the first node on the first image for labels classifying the first node and the corresponding node on the second image as normal, wherein the unary potential is derived by comparing features associated with the first node and the corresponding node on the second image and penalizing a label assigned to the first node for dissimilarity between these features;

c. determining a binary potential for a pair of neighboring nodes in the first image wherein determining the binary potential for the pair of neighboring nodes and a pair of combined labels applied to the pair of neighboring nodes comprises penalizing the pair of combined labels applied to the neighboring nodes for dissimilarity in the mapping parts of the pair of labels and penalizing the pair of labels applied to the pair of neighboring nodes for dissimilarity in the classification parts of the pair of labels;

d. inputting results derived from the calculating the unary and binary potentials into a probabilistic model, wherein the probabilistic model is a Conditional Random Field model; and

e. computing the probability of cancer in each node by estimating Posterior Marginal Probability for each label applied to each node in the framework of the Conditional Random Field model and then marginalizing over all possible mappings of the node and over labels of its corresponding nodes; and

providing the probable cancer regions.

14. The computer storage media of claim 13, wherein the two mammogram images comprise a first mammogram image of a breast and a second mammogram image comprising one of an image of the breast taken at a later date or an image of an opposite breast.