Methods for microcalcification detection of breast cancer on digital tomosynthesis mammograms

• US 8,977,019 B2
• Filed: 02/11/2011
• Issued: 03/10/2015
• Est. Priority Date: 02/11/2010
• Status: Active Grant

• Current Assignee: Board of Regents of the University of Michigan (University of Michigan), University of Michigan
• Original Assignee: Board of Regents of the University of Michigan (University of Michigan)
• Inventors: Chan, Heang-Ping, Sahiner, Berkman, Hadjiiski, Lubomir M., Wei, Jun, Helvie, Mark A.
• Primary Examiner(s): Chang, Jon
• Assistant Examiner(s): Mistry, Oneal R
• Application Number:
  US13/514,804
• Publication Number:
  US 20120294502A1
• Time in Patent Office: 1,488 Days
• Field of Search: None
• US Class Current: 382/128
• CPC Class Codes: A61B 6/025 Tomosynthesis
• CPC Class Codes: A61B 6/502 for diagnosis of breast, i.e. mammography
• CPC Class Codes: A61B 6/5211 involving processing of medical diagnostic data
• CPC Class Codes: G06T 2207/10112 Digital tomosynthesis [DTS]
• CPC Class Codes: G06T 2207/30068 Mammography; Breast
• CPC Class Codes: G06T 7/0012 Biomedical image inspection

Abstract

A computer-aided detection system to detect clustered microcalcifications in digital breast tomosynthesis (DBT) is disclosed. The system performs detection in 2D images and a reconstructed 3D volume. The system may include an initial prescreening of potential microcalcifications by using one or more 3D calcification response function (CRF) values modulated by an enhancement method to identify high response locations in the DBT volume as potential signals. Microcalcifications may be enhanced using a Multi-Channel Enhancement method. Locations detected using these methods can be identified and the potential microcalcifications may be extracted. The system may include object segmentation that uses region growing guided by the enhancement-modulated CRF values, gray level voxel values relative to a local background level, or the original DBT voxel values. False positives may be reduced by descriptors of characteristics of microcalcifications. Detected locations of clusters and a cluster significance rating of each cluster may be output and displayed.

Claims

1. A computer-implemented method for detecting a microcalcification cluster in reconstructed volumes of a three-dimensional digital breast tomosynthesis image, comprising:

receiving, at a computing device, a three-dimensional digital breast tomosynthesis image;

reconstructing, at the computing device, the three-dimensional digital breast tomosynthesis image to create a reconstructed three-dimensional digital breast tomosynthesis image;

enhancing, at the computing device, a microcalcification candidate in each slice of the reconstructed three-dimensional digital breast tomosynthesis image to obtain an enhanced three-dimensional digital breast tomosynthesis image by simultaneously;

1) determining a multi-scale calcification response for each voxel (x,y,z) of the microcalcification candidate, and

2) one of;

a) enhancing a signal-to-noise ratio of a microcalcification in the reconstructed three-dimensional digital breast tomosynthesis image, or

b) enhancing a signal intensity of the microcalcification in the reconstructed three-dimensional digital breast tomosynthesis image;

wherein the multi-scale calcification response includes;

a) filtering the reconstructed three-dimensional digital breast tomosynthesis image with a plurality of smoothing filters where an amount of smoothing for each smoothing filter is determined by a scale, and the plurality of filters includes a set of scales;

b) finding a calcification response at each voxel of the reconstructed three-dimensional digital breast tomosynthesis image for each scale of the set of scales, where, among the set of scales, the calcification response at a central voxel of a three-dimensional microcalcification is higher when the scale is such that a shape of the smoothing filter approximately matches a shape of the three-dimensional microcalcification; and

c) combining, among the set of scales, each calcification response using a non-linear operator at each voxel of the reconstructed three-dimensional digital breast tomosynthesis image; and

detecting, at the computing device, potential individual microcalcifications that form a microcalcification cluster within the enhanced three-dimensional digital breast tomosynthesis image.

2. The computer-implemented method of claim 1, wherein detecting the seed object includes weighting a multi-scale calcification response volume for each voxel (x,y,z) by either:

a) the signal-to-noise ratio enhanced three-dimensional digital breast tomosynthesis image or b) a signal intensity enhanced three-dimensional digital breast tomosynthesis image.

3. The computer-implemented method of claim 2, wherein the weighting creates an enhancement-modulated calcification response function at each voxel of the enhanced three-dimensional digital breast tomosynthesis image and the detected seed object includes a voxel having a value of the enhancement-modulated calcification response function that exceeds a threshold.

4. The computer-implemented method of claim 1, wherein detecting potential individual microcalcifications based on the location of the seed object includes:

identifying a plurality of voxels within either;

a) the signal-to-noise ratio enhanced three-dimensional digital breast tomosynthesis image or b) the signal intensity enhanced three-dimensional digital breast tomosynthesis image, wherein the identified plurality of voxels have a non-zero calcification response function; and

iteratively identifying a subset of the plurality of identified voxels, wherein a signal-to-noise ratio value of each member of the subset is above a threshold.

5. The computer-implemented method of claim 1, wherein enhancing the signal-to-noise ratio of the microcalcification or enhancing the signal intensity of the microcalcification creates a three-dimensional enhancement-modulated calcification response image.

6. The computer-implemented method of claim 5, wherein determining the multi-scale calcification response for each voxel (x,y,z) of the microcalcification candidate includes, for the set of scales {σ

} surrounding the microcalcification candidate, the multi-scale calcification response at a voxel (x,y,z) defined as;

7. A non-transitory computer-readable medium having instructions stored thereon, the instructions when executed by a processor detect a microcalcification cluster in reconstructed volumes of a three-dimensional digital breast tomosynthesis image, causing the processor to:

receive a three-dimensional digital breast tomosynthesis image;

reconstruct the three-dimensional digital breast tomosynthesis image to create a reconstructed three-dimensional digital breast tomosynthesis image;

enhance a microcalcification candidate in each slice of the reconstructed three-dimensional digital breast tomosynthesis image to obtain a three-dimensional enhancement-modulated calcification response image;

detect a seed object in the three-dimensional enhancement-modulated calcification response image, wherein the seed object includes a plurality of voxels; and

detect a plurality of microcalcification candidates based on criteria including one or more of a proximity of each microcalcification candidate to the seed object and a signal-to-noise ratio value of each microcalcification candidate exceeding a threshold;

wherein the instruction to enhance the microcalcification candidate in each slice of the three-dimensional digital breast tomosynthesis image to obtain the three-dimensional enhancement-modulated calcification response image includes a multi-scale calcification response function to;

construct a Hessian matrix for each voxel of the reconstructed three-dimensional digital breast tomosynthesis image;

select a voxel of the reconstructed three-dimensional digital breast tomosynthesis image wherein all eigenvalues of the Hessian matrix corresponding to the selected voxel are negative;

determine a ratio between a square of a smallest magnitude eigenvalue and a negative value of a largest magnitude eigenvalue, wherein the ratio includes a calcification response;

create an enhanced three-dimensional digital breast tomosynthesis image by enhancing, for each voxel of the three-dimensional digital breast tomosynthesis image, either;

a) a signal to noise ratio to obtain the enhanced three-dimensional digital breast tomosynthesis image or b) a signal intensity value to obtain the enhanced three-dimensional digital breast tomosynthesis image; and

weight each voxel of a multi-scale calcification response volume by a value of either;

a) the signal-to-noise ratio enhanced three-dimensional digital breast tomosynthesis image or b) the signal intensity value enhanced three-dimensional digital breast tomosynthesis image to obtain the three-dimensional enhancement-modulated calcification response image from the three-dimensional digital breast tomosynthesis image.

8. A computer system for detecting a microcalcification cluster in reconstructed volumes of a three-dimensional digital breast tomosynthesis image, the system comprising:

a processor;

a memory;

a reconstruction module stored in the memory and executed by the processor to reconstruct the three-dimensional digital breast tomosynthesis image to create a reconstructed three-dimensional digital breast tomosynthesis image;

an enhancement module stored in the memory and executed by the processor to enhance a microcalcification candidate in each slice of the reconstructed three-dimensional digital breast tomosynthesis image to obtain a three-dimensional enhancement-modulated calcification response image;

a seed detection module stored in the memory and executed by the processor to detect a seed object in the three-dimensional enhancement-modulated calcification response image, wherein the seed object includes a plurality of voxels;

a microcalcification candidate detection module stored in the memory and executed by the processor to detect a plurality of microcalcification candidates based on criteria including one or more of a proximity of each microcalcification candidate to the seed object and a signal-to-noise ratio value of each microcalcification candidate exceeding a threshold;

a multi-scale calcification response module stored in the memory and executed by the processor to obtain a multi-scale calcification response volume from a reconstructed volume of the three-dimensional digital breast tomosynthesis image by applying a multi-scale calcification response function to each voxel of the reconstructed volume, wherein the multi-scale calcification response function is configured to;

smooth the enhanced reconstructed three-dimensional digital breast tomosynthesis image;

construct a Hessian matrix for each voxel of the enhanced three-dimensional digital breast tomosynthesis image;

select a voxel of the enhanced three-dimensional digital breast tomosynthesis image wherein all eigenvalues of the Hessian matrix corresponding to the selected voxel are negative; and

determine a ratio between a square of a smallest magnitude eigenvalue and a negative value of a largest magnitude eigenvalue, wherein the ratio includes a calcification response;

create an enhanced three-dimensional digital breast tomosynthesis image by enhancing, for each voxel of the three-dimensional digital breast tomosynthesis image, either;

a) a signal to noise ratio to obtain the enhanced three-dimensional digital breast tomosynthesis image or b) a signal intensity value to obtain the enhanced three-dimensional digital breast tomosynthesis image; and

a weighting module stored in the memory and executed by the processor to weight each voxel of the three-dimensional multi-scale calcification response volume by a value of either;

a) the signal-to-noise ratio enhanced three-dimensional digital breast tomosynthesis image or b) the signal intensity value enhanced three-dimensional digital breast tomosynthesis image to obtain the three-dimensional enhancement-modulated calcification response image from the three-dimensional digital breast tomosynthesis image.

9. The computer system of claim 8, wherein the weighting module is further executed by the processor to weight the multi-scale calcification response volume by the signal-to-noise enhanced volume according to an equation:

EMCR(x, y, z)=E(x, y, z)*[I(x, y, z){circle around (×

)}F(x, y, z)] where I(x, y, z){circle around (×

)}F(x, y, z)

denotes the signal-to-noise enhanced volume and E(x, y, z) denotes the multi-scale calcification response volume.

10. A computer-implemented method for determining a strength of a microcalcification candidate in an image for differentiation of true and false microcalcifications comprising:

receiving an image including one of a two-dimensional projection view or a three-dimensional digital breast tomosynthesis slice image;

identifying a region of interest within the image including a plurality of microcalcification candidates;

characterizing the region of interest as a vector g including a set of channel response {g₁, ... , g_N} given a multi-channel set;

differentiating each of the plurality of microcalcification candidates as one of a true microcalcification or a false microcalcification according to a linear classification model;

D(g)=( m₂−

m₁)^TΣ

<sup>−

1</sup>g

wherein m_kincludes a mean vector for class k where k=1,2, Σ

includes a covariance matrix estimated from training samples of true microcalcifications and false microcalcifications, and D(g) includes a multi-channel enhancement (MCE) response representing the strength of each of the plurality of microcalcification candidates.

11. The computer-implemented method of claim 10, comprising using the MCE response to differentiate true and false microcalcifications and generating a measure of the microcalcification cluster including a cluster likelihood score (CLS), a level of suspicion (LOS), and an overall cluster significance rating (CSR).

12. The computer-implemented method of claim 11, further comprising estimating the CSR for each plurality of microcalcification candidates by:

extracting a plurality of features from the three-dimensional digital breast tomosynthesis image;

extracting the MCE response for each plurality of microcalcification candidates;

combining the extracted features and the MCE response into a CLS score by using a classifier trained to classify the detected plurality of microcalcification candidates as one of true or false;

combining the extracted features and the MCE response into a LOS score by using a classifier trained to classify the detected plurality of microcalcification candidates as one of malignant or benign;

weighting the CLS value by the LOS score.

13. A non-transitory computer-readable medium having instructions stored thereon, the instructions when executed by a processor determine a strength of a microcalcification candidate in an image for differentiation of true and false microcalcifications, causing the processor to:

receive an image including one of a two-dimensional projection view or a three-dimensional digital breast tomosynthesis slice image;

identify a region of interest within the image including a plurality of microcalcification candidates;

characterize the region of interest as a vector g including a set of channel response {g₁, ... , g_N} given a multi-channel set;

differentiate each microcalcification candidate as one of a true microcalcification or a false microcalcification according to a linear classification model;

D(g)=( m₂−

m₁)^TΣ

<sup>−

1</sup>g

wherein m_k includes a mean vector for class k where k=1,2, Σ

includes a covariance matrix estimated from training samples of true microcalcifications and false microcalcifications, and D(g) includes a multi-channel enhancement (MCE) response representing the strength of each microcalcification candidate.

14. The non-transitory computer-readable medium of claim 13, further comprising instructions that, when executed by a processor, cause the processor to estimate a cluster significance rating (CSR) for each microcalcification candidate by:

extracting a plurality of features from the three-dimensional digital breast tomosynthesis image;

extracting the MCE response for each microcalcification candidate;

combining the extracted features and the MCE response into a CLS score by using a classifier trained to classify the detected plurality of microcalcification candidates as true or false;

combining the extracted features and the MCE response into a LOS score by using a classifier trained to classify the detected plurality of microcalcification candidates as malignant or benign;

weighting the CLS value by the LOS.

15. A computer system for determining a strength of a microcalcification candidate in an image for differentiation of true and false microcalcifications, the system comprising:

a processor;

a memory;

a receiving module stored in the memory and executed by the processor to receive an image including one of a two-dimensional projection view or a three-dimensional digital breast tomosynthesis slice image;

a prescreening module stored in the memory and executed by the processor to identify a region of interest within the image including a plurality of microcalcification candidates;

a vector module stored in the memory and executed by the processor to characterize the region of interest as a vector g including a set of channel response {g₁, ... , g_N} given a multi-channel set;

a differentiation module stored in the memory and executed by the processor to differentiate each microcalcification candidate as one of a true microcalcification or a false microcalcification according to a linear classification model;

D(g) =( m₂−

m₁)^TΣ

<sup>−

1</sup>g

wherein m_kincludes a mean vector for class k where k=1,2, Σ

includes a covariance matrix estimated from training samples of true microcalcifications and false microcalcifications, and D(g) includes a multi-channel enhancement (MCE) response representing the strength of each microcalcification candidate.

16. The computer system of claim 15, further comprising a cluster significance rating (CSR) estimation module stored in the memory and executed by the processor to estimate the CSR for each microcalcification candidate by:

extracting a plurality of features from the three-dimensional digital breast tomosynthesis image;

extracting the MCE response for each microcalcification candidate;

combining the extracted features and the MCE response into a CLS score by using a classifier trained to classify the detected plurality of microcalcification candidates as true or false;

combining the extracted features and the MCE response into a LOS score by using a classifier trained to classify the detected plurality of microcalcification candidates as malignant or benign;

weighting the CLS value by the LOS.

17. A computer-implemented method for detecting a microcalcification cluster in an image, comprising:

enhancing, at a computing device, a received image, the image including one of a two-dimensional projection view or a three-dimensional digital tomosynthesis slice image and a plurality of microcalcification candidates, the enhancing including differentiating each of the plurality of microcalcification candidates as one of a true microcalcification or a false microcalcification according to a linear classification model;

D(g)=( m₂−

m₁)^TΣ

<sup>−

1</sup>g

wherein m_k includes a mean vector for class k where k=1,2, Σ

includes a covariance matrix estimated from training samples of true microcalcifications and false microcalcifications, and D(g) includes a multi-channel enhancement (MCE) response representing a strength of each of the plurality of microcalcification candidates;

back-projecting, at the computing device, an MCE-response for each microcalcification candidate on the two-dimensional projection view to the three-dimensional tomosynthesis image

determining, at the computing device, a multi-scale calcification response function for the three-dimensional tomosynthesis image by;

smoothing the enhanced reconstructed three-dimensional digital breast tomosynthesis image;

constructing a Hessian matrix for each voxel of the enhanced reconstructed three-dimensional digital breast tomosynthesis image;

selecting a voxel of the enhanced reconstructed three-dimensional digital breast tomosynthesis image wherein all eigenvalues of the Hessian matrix corresponding to the selected voxel are negative; and

determining a ratio between a square of a smallest magnitude eigenvalue and a negative value of a largest magnitude eigenvalue, wherein the ratio includes a calcification response;

weighting, at the computing device, the multi-scale calcification response volume by the back-projected MCE-response;

detecting, at the computing device, a seed object in the weighted multi-scale calcification response volume.

18. A non-transitory computer-readable medium having non-transitory instructions stored thereon, the instructions when executed by a processor detect a microcalcification cluster in an image, causing the processor to:

enhance a received image, the image including one of a two-dimensional projection view or a two-dimensional digital tomosynthesis slice image and a plurality of microcalcification candidates, the enhance instruction further causing the processor to differentiate each of the plurality of microcalcification candidates as one of a true microcalcification or a false microcalcification according to a linear classification model;

D(g)=( m₂−

m₁)^TΣ

⁻

g

wherein includes a mean vector for class k where k=1,2, includes a covariance matrix estimated from training samples of true microcalcifications and false microcalcifications, and D(g) includes a multi-channel enhancement (MCE) response representing a strength of each of the plurality of microcalcification candidates;

back-project an MCE-response for each microcalcification candidate on the two-dimensional projection view to the three-dimensional tomosynthesis image

determine a multi-scale calcification response function for the three-dimensional tomosynthesis image by;

smoothing the enhanced reconstructed three-dimensional digital breast tomosynthesis image;

constructing a Hessian matrix for each voxel of the enhanced reconstructed three-dimensional digital breast tomosynthesis image;

selecting a voxel of the enhanced reconstructed three-dimensional digital breast tomosynthesis image wherein all eigenvalues of the Hessian matrix corresponding to the selected voxel are negative; and

determining a ratio between a square of a smallest magnitude eigenvalue and a negative value of a largest magnitude eigenvalue, wherein the ratio includes a calcification response;

weight the multi-scale calcification response volume by the back-projected MCE-response;

detect a seed object in the weighted multi-scale calcification response volume, wherein the seed object includes a plurality of voxels.

19. A computer system for detecting a microcalcification cluster in an image, the system comprising:

a processor;

a memory;

a two-dimensional image enhancement module stored in the memory and executed by the processor to enhance a received image, the image including one of a two-dimensional projection view or a two-dimensional digital tomosynthesis slice image and a plurality of microcalcification candidates, the two-dimensional image enhancement module further differentiating each of the plurality of microcalcification candidates as one of a true microcalcification or a false microcalcification according to a linear classification model;

D(g)=( m₂−

m₁)^TΣ

<sup>−

1</sup>g

wherein m_k includes a mean vector for class k where k=1,2, Σ

includes a covariance matrix estimated from training samples of true microcalcifications and false microcalcifications, and D(g) includes a multi-channel enhancement (MCE) response representing a strength of each of the plurality of microcalcification candidates;

a seed object detection module stored in the memory and executed by the processor to;

back-project an MCE-response for each microcalcification candidate on the two-dimensional projection view to the three-dimensional tomosynthesis image

determine a multi-scale calcification response function for the three-dimensional tomosynthesis image by;

smoothing the enhanced reconstructed three-dimensional digital breast tomosynthesis image;

constructing a Hessian matrix for each voxel of the enhanced reconstructed three-dimensional digital breast tomosynthesis image;

selecting a voxel of the enhanced reconstructed three-dimensional digital breast tomosynthesis image wherein all eigenvalues of the Hessian matrix corresponding to the selected voxel are negative; and

determining a ratio between a square of a smallest magnitude eigenvalue and a negative value of a largest magnitude eigenvalue, wherein the ratio includes a calcification response;

weight the multi-scale calcification response volume by the back-projected MCE-response; and

detect a seed object in the weighted multi-scale calcification response volume

a clustering module stored in the memory and executed by the processor to detect a microcalcification cluster based on a location of the seed object in the weighted multi-scale calcification response volume, the

microcalcification cluster having a plurality of microcalcification candidates with a calcification response above a threshold.