COMPUTER-AIDED DETECTION OF ANATOMICAL ABNORMALITIES IN X-RAY TOMOSYNTHESIS IMAGES

• US 20100246913A1
• Filed: 03/31/2009
• Published: 09/30/2010
• Est. Priority Date: 03/31/2009
• Status: Active Grant

• Current Assignee: Hologic Incorporated
• Original Assignee: Hologic Incorporated
• Inventors: MARSHALL, Julian, ZHANG, Xiangwei, SRINIVAS, Chukka, KREEGER, Kevin A., CHUI, Haili, ZHANG, Wei
• Granted Patent: US 8,223,916 B2
• Time in Patent Office: Days
• US Class Current: 382/131
• CPC Class Codes: A61B 6/025 Tomosynthesis
• CPC Class Codes: A61B 6/502 for diagnosis of breast, i.e. mammography
• CPC Class Codes: G06T 11/006 Inverse problem, transformation from projection-space into object-space, e.g. transform methods, back-projection, algebraic methods
• CPC Class Codes: G06T 2207/10112 Digital tomosynthesis [DTS]
• CPC Class Codes: G06T 2207/20068 Projection on vertical or horizontal image axis
• CPC Class Codes: G06T 2207/20212 Image combination
• CPC Class Codes: G06T 2207/30068 Mammography; Breast
• CPC Class Codes: G06T 2211/436 Limited angle
• CPC Class Codes: G06T 7/0012 Biomedical image inspection
• CPC Class Codes: G06V 10/462 Salient features, e.g. scale invariant feature transforms [SIFT]

Abstract

Methods, systems, and related computer program products for computer-aided detection (CAD) of anatomical abnormalities in a breast volume based on a plurality of two-dimensional x-ray tomosynthesis projection images thereof is described. Each projection image is processed according to at least one predetermined feature extraction algorithm to generate at least one projection feature array corresponding thereto. For each of the at least one predetermined features extracted, the plurality of corresponding projection feature arrays is backprojected according to a predetermined tomosynthesis reconstruction algorithm to form a plurality of two-dimensional tomosynthesis reconstructed feature arrays. Each pixel in the three-dimensional breast geometry is then individually classified as being either a pixel of interest, or not a pixel of interest, based upon the at least one tomosynthesis reconstructed feature array value corresponding to that pixel. CAD detections for the breast volume are then generated based on the identified pixels of interest.

Claims

1. A method for computer-aided detection (CAD) of anatomical abnormalities in a breast volume based on a plurality of two-dimensional x-ray tomosynthesis projection images thereof acquired at a respective plurality of x-ray tomosynthesis projection angles, comprising:

for each of the plurality of projection images, processing that projection image separately from the other projection images according to at least one predetermined feature extraction algorithm to generate at least one projection feature array corresponding thereto;

for each of the at least one predetermined features extracted, backprojecting the plurality of corresponding projection feature arrays according to a predetermined tomosynthesis reconstruction algorithm to form a plurality of two-dimensional tomosynthesis reconstructed feature arrays, the imaged breast volume having a three-dimensional geometry characterized by a number of levels and a number of pixels per level, wherein said plurality of two-dimensional tomosynthesis reconstructed feature arrays collectively occupy said three-dimensional geometry;

individually classifying each pixel in each level of said three-dimensional geometry as being one of (i) a pixel of interest and (ii) not a pixel of interest based upon the at least one tomosynthesis reconstructed feature array value corresponding to that pixel location; and

generating CAD detections for the breast volume based on the identified pixels of interest.

2. The method of claim 1, wherein said at least one predetermined feature extraction algorithm performed on each said projection image comprises a quantum noise-compensating microcalcification enhancement algorithm designed to generate a projection quantum noise-compensated microcalcification feature array.

3. The method of claim 2 wherein said quantum noise-compensating microcalcification enhancement algorithm comprises, for an n^th of said projection images, the steps of:

generating a microcalcification contrast-enhanced version MCEF_n(x,y) of the n^th projection image using a high-pass convolution kernel sized according to an expected microcalcification size range;

generating a smoothed version SF_n(x,y) of the n^th projection image using a smoothing filter that is relatively large compared to the expected microcalcification size range;

generating a lookup table LUT_n(pixval) that, for an argument value pixval, corresponds to a statistical variation measure of MCEF_n(x,y) for those pixels in the n^th projection image for which SF_n(x,y) is equal to the pixval; and

setting said projection quantum noise-compensated microcalcification feature array equal to MCEF_n(x,y)/SF_n(x,y) for each pixel (x,y)_n in the n^th projection image.

4. The method of claim 3, said backprojection forming a plurality of tomosynthesis reconstructed quantum noise-compensated microcalcification feature arrays, wherein said individually classifying each pixel in each level of said three-dimensional geometry comprises applying a global threshold value to each pixel in said plurality of tomosynthesis reconstructed quantum noise-compensated microcalcification feature arrays.

5. The method of claim 3, wherein said generating CAD detections for the breast volume based on the identified pixels of interest comprises grouping adjacent ones of said pixels of interest into candidate microcalcification spots, clustering said candidate microcalcification spots, computing a plurality of features characterizing said microcalcification spots/clusters, and classifying said spots/clusters as being CAD detections, or not being CAD detections, based on said computed features.

6. The method of claim 1, wherein said individually classifying each pixel as being one of (i) a pixel of interest and (ii) not a pixel of interest consists of a non-neighborhood process based solely on the least one tomosynthesis reconstructed feature array value at that pixel location.

7. The method of claim 6, wherein said three-dimensional geometry occupied by said plurality of two-dimensional tomosynthesis reconstructed feature arrays spans substantially an entirety of the imaged breast volume at a diagnostically sufficient resolution for suspicious abnormality detection, whereby said pixels of interest are identified without requiring either of (a) prior reconstruction of a tomosynthesis reconstructed image volume from said plurality of projection images, or (b) neighborhood-based processing of the two-dimensional tomosynthesis reconstructed feature arrays.

8. The method of claim 1, wherein said generating CAD detections for the breast volume based on the identified pixels of interest comprises discarding from consideration all pixels classified as not being of interest.

9. The method of claim 1, further comprising determining an outline of the breast volume based on a two-dimensional extent of breast tissue appearing in one or more of said projection images, wherein said three-dimensional breast geometry is based on said determined outline.

10. The method of claim 1, further comprising backprojecting the plurality of projection image arrays according to said predetermined tomosynthesis reconstruction algorithm to form a plurality of two-dimensional tomosynthesis reconstructed images, wherein said individually classifying each pixel in each level of said three-dimensional geometry is further based upon a pixel value of the tomosynthesis reconstructed image at that pixel location and level.

11. The method of claim 1, wherein said at least one predetermined feature extraction algorithm is configured to generate, for an n^th of said projection images, each of a projection mass density metric array G1p_n(x,y), a projection mass isotropy metric array G2p_n(x,y), a projection stellateness magnitude metric array F1p_n(x,y), and a projection stellateness isotropy metric array F2p_n(x,y).

12. The method of claim 11, said backprojection forming, for an m^th level of said tomosynthesis reconstruction geometry, a tomosynthesis reconstructed mass density metric array G1r_m(x,y), a tomosynthesis reconstructed mass isotropy metric array G2r_m(x,y), a tomosynthesis reconstructed stellateness magnitude metric array F1r_m(x,y), and a tomosynthesis reconstructed stellateness isotropy metric array F2r_m(x,y), wherein said individually classifying each pixel for said m^th level in said three-dimensional geometry comprises applying said G1r_m(x,y), G2r_m(x,y), F1r_m(x,y), and F2r_m(x,y) features to a predetermined classifier algorithm.

13. The method of claim 12, wherein said predetermined classifier algorithm is selected from the group consisting of:

artificial neural networks, rule-based classifiers, decision trees, and support vector machines.

14. The method of claim 12, wherein said generating CAD detections for the breast volume based on the identified pixels of interest comprises grouping said pixels of interest into candidate masses, computing a plurality of features characterizing said candidate masses, and classifying said masses as being CAD detections, or not being CAD detections, based on said computed features.

15. A system for detecting suspicious microcalcifications in a breast volume based on a plurality of two-dimensional x-ray tomosynthesis projection images thereof acquired at a respective plurality of x-ray tomosynthesis projection angles, the system comprising a processor programmed to perform the steps of:

jointly processing the plurality of projection images to generate a noise compensation function based on joint noise statistics in the plurality of projection images;

processing each of said plurality of projection images using a microcalcification-enhancing filter, a spatial smoothing filter, and said noise compensation function to generate a respective plurality of projection noise-compensated microcalcification feature arrays;

backprojecting said plurality of projection noise-compensated microcalcification feature arrays according to a predetermined tomosynthesis reconstruction algorithm to form a plurality of two-dimensional tomosynthesis reconstructed feature arrays, the imaged breast volume having a three-dimensional geometry characterized by a number of levels and a number of pixels per level, wherein said plurality of two-dimensional tomosynthesis reconstructed feature arrays collectively occupy said three-dimensional geometry;

individually classifying each pixel in each level of said three-dimensional geometry as being one of (i) a pixel of interest and (ii) not a pixel of interest based upon the at least one tomosynthesis reconstructed feature array value corresponding to that pixel location; and

processing the identified pixels of interest to detect suspicious microcalcifications in the breast volume.

16. The system of claim 15, said plurality of projection images having “

N”

members, wherein said jointly processing said N projection images to generate said noise compensation function comprises;

for each n^th projection image, n=1 . . . N, generating a microcalcification contrast-enhanced version MCEF_n(x,y) thereof using a high-pass convolution kernel sized according to an expected microcalcification size range;

for each n^th projection image, n=1 . . . N, generating a smoothed version SF_n(x,y) thereof using a smoothing filter that is relatively large compared to the expected microcalcification size range; and

generating said noise compensation function as a lookup table LUT (pixval) that, for each argument value pixval, corresponds to a statistical variation in the values of MCEF_n(x,y) for all pixels in the N projection images having corresponding values of SF_n(x,y) that are equal to pixval.

17. The system of claim 16, wherein for each pixel (x,y)_n in each n^th projection image, said noise-compensated microcalcification feature arrays is assigned the value of MCEF_n(x,y)/LUT(SF_n(x,y)).

18. The system of claim 17, wherein said individually classifying each pixel as being one of (i) a pixel of interest and (ii) not a pixel of interest consists of a non-neighborhood process based solely on the tomosynthesis reconstructed feature array value at that pixel location, and wherein said three-dimensional geometry occupied by said plurality of two-dimensional tomosynthesis reconstructed feature arrays spans substantially an entirety of the imaged breast volume at a diagnostically sufficient resolution for microcalcification detection, whereby said pixels of interest are identified without requiring either of (a) prior reconstruction of a tomosynthesis reconstructed image volume from said plurality of projection images, or (b) neighborhood-based processing of the two-dimensional tomosynthesis reconstructed feature arrays.

19. A computer readable medium embodying a computer program product for computer-aided detection (CAD) of anatomical abnormalities in a breast volume based on a plurality of two-dimensional x-ray tomosynthesis projection images thereof acquired at a respective plurality of x-ray tomosynthesis projection angles, comprising:

computer code for processing each of the plurality of projection images according to at least one predetermined feature extraction algorithm to generate at least one projection feature array corresponding thereto;

computer code for backprojecting, for each of the at least one predetermined features extracted, the plurality of corresponding projection feature arrays according to a predetermined tomosynthesis reconstruction algorithm to form a plurality of two-dimensional tomosynthesis reconstructed feature arrays, the imaged breast volume having a three-dimensional geometry characterized by a number of levels and a number of pixels per level, wherein said plurality of two-dimensional tomosynthesis reconstructed feature arrays collectively occupy said three-dimensional geometry;

computer code for individually classifying each pixel in each level of said three-dimensional geometry as being one of (i) a pixel of interest and (ii) not a pixel of interest based upon the at least one tomosynthesis reconstructed feature array value corresponding to that pixel location; and

computer code for generating CAD detections for the breast volume based on the identified pixels of interest.

20. A computer readable medium of claim 19, wherein said predetermined tomosynthesis reconstruction algorithm used in said computer code for backprojecting is selected from the group consisting of:

simple backprojection, filtered backprojection, cone-beam filtered backprojection, order-statistics based backprojection, and matrix inversion tomosynthesis.