Method and system for the detection of microcalcifications in digital mammograms

• US 5,491,627 A
• Filed: 05/13/1993
• Issued: 02/13/1996
• Est. Priority Date: 05/13/1993
• Status: Expired due to Term

• Current Assignee: ARCH Development Corporation (University of Chicago)
• Original Assignee: ARCH Development Corporation (University of Chicago)
• Inventors: Zhang, Wei, Doi, Kunio
• Primary Examiner(s): Harvey, Jack B.
• Assistant Examiner(s): CHUNG TRANS, XUONG MY
• Application Number:
  US08/060,531
• Time in Patent Office: 1,006 Days
• Field of Search: 364/413.13, 364/413.22, 382/2, 382/6, 382/19, 128/664
• US Class Current: 600/408
• CPC Class Codes: G06T 2207/10116 X-ray image
• CPC Class Codes: G06T 2207/20084 Artificial neural networks [ANN]
• CPC Class Codes: G06T 2207/30068 Mammography; Breast
• CPC Class Codes: G06T 7/0012 Biomedical image inspection
• CPC Class Codes: G16H 30/20 for handling medical images, e.g. DICOM, HL7 or PACS
• CPC Class Codes: G16H 50/20 for computer-aided diagnosis, e.g. based on medical expert systems
• CPC Class Codes: Y10S 128/924 using artificial intelligence
• CPC Class Codes: Y10S 128/925 Neural network

Abstract

A method and system for the detection of microcalcifications in digital mammograms. Digital mammograms are obtained and regions-of-interest (ROIs) are selected therefrom which contain suspected microcalcifications, either individual or clustered microcalcifications. The suspect ROIs are background-trend corrected, followed by Fourier transformation and power spectrum calculation to perform detection in the frequency domain. Detection can also be carried out in the spatial domain by omitting the Fourier transformation and power spectrum calculation. The ROI is then scaled for input into a neural network trained to detect microcalcifications. The neural network outputs ROIs with detected microcalcifications. The method and system can also include normalizing the background-trend corrected ROIs and imputing the normalized ROI to a shift-invariant neural network trained to detect microcalcifications. The output ROI of the shift-invariant neural network is thresholded to remove additional false positive detections, and then the thresholded ROI undergoes a cluster detection to detect clustered microcalcifications. Feature extraction techniques can be applied to the remaining ROI to remove additional false positive detections.

Claims

1. A method of detecting microcalcifications in a digital mammogram, comprising:

obtaining a digital mammogram;

extracting regions of interest from said digital mammogram suspected of containing a microcalcification;

converting said regions of interest into corresponding numerical data;

inputting said numerical data to a shift-invariant neural network trained to detect microcalcifications for processing of said numerical data by said neural network, said neural network outputting corresponding output images;

detecting a microcalcification in said digital mammogram using said output images.

2. A method as recited in claim 1, wherein extracting regions of interest comprises extracting regions of interest suspected of containing a clustered microcalcification.

3. A method as recited in claim 2, wherein said extracting of regions of interest suspected of containing clustered microcalcifications includes selecting regions of interest containing false-positive microcalcifications;

and

wherein processing by said neural network includes removing a portion of said false-positive microcalcifications selected in said extracting step.

4. A method as recited in claim 3, wherein extracting a region of interest comprises extracting a region of interest suspected of containing an individual microcalcification.

5. A method as recited in claim 1, wherein processing said region of interest comprises:

background-trend correcting said regions of interest to produce background-trend corrected regions of interest; and

converting said background-trend corrected region of interest into numerical data.

6. A method as recited in claim 1, wherein said extracting regions of interest suspected of containing a microcalcification includes selecting regions of interest containing false-positive microcalcifications;

and

wherein processing by said neural network includes removing a portion of said false-positive microcalcifications selected in said extracting step.

7. A method as recited in claim 1, wherein extracting said regions of interest comprises extracting spatial domain regions of interest.

8. A system for detecting microcalcifications in a digital mammogram, comprising:

a device for obtaining a digital mammogram;

a detector connected to said device which detects suspected microcalcifications in said digital mammogram;

a region of interest selector, connected to said device and said detector, which selects regions of interest in said digital mammogram corresponding to said suspected microcalcifications;

a region of interest processing device outputting processed regions of interest;

an input data scaler connected to said region of interest processing device for numerically scaling said processed regions of interest;

a shift-invariant neural network receiving as input data said numerically scaled regions of interest from said input data scaler and outputting corresponding output images; and

a microcalcification detector for detecting microcalcifications in said output images.

9. A system as recited in claim 8, wherein said region of interest extractor comprises a spatial domain region of interest extractor.

10. A method of detecting clustered microcalcifications in a digital mammogram, comprising:

obtaining a digital mammogram;

extracting regions of interest from said digital mammogram suspected of containing a clustered microcalcification;

converting said regions of interest into corresponding numerical data;

inputting said numerical data into a shift-invariant neural network trained to detect clustered microcalcifications;

processing said numerical data by said shift-invariant neural network to produce corresponding output images;

detecting a clustered microcalcification in said digital mammogram using said output images.

11. A method as recited in claim 10, wherein said extracting regions of interest suspected of containing a microcalcification includes selecting regions of interest containing false-positive microcalcifications using a computer-aided diagnostic technique;

and

wherein processing by said shift-invariant neural network includes removing a portion of said false-positive microcalcifications selected in said extracting step.

12. A method as recited in claim 10, wherein extracting said regions of interest comprises:

detecting suspected clustered microcalcifications in said digital mammogram using a computer-aided diagnostic technique;

selecting regions of interest from said digital mammogram corresponding to said suspected clustered microcalcifications.

13. A method as recited in claim 10, further comprising:

processing said input data using a neural network having an input layer having a plurality of input units, at least one hidden layer having a plurality of groups, each of said groups having a plurality of units, an output layer, and connection weights between said layers; and

selecting said connection weights to be shift-invariant as

space="preserve" listing-type="equation">w(i,j;

x,y)=w(i-x,j-y)

where w(i,j;

x,y) denotes a connection weight between a unit at a location (i,j) in a first layer and a unit at a location (x,y) in a subsequent layer.

14. A method as recited in claim 13, further comprising:

training said neural network using a modified error back propagation method;

##EQU7## and in the case of said output layer,

space="preserve" listing-type="equation">δ

.sub.l.sup.L (x,y)=ƒ

'"'"'(N.sub.l.sup.L (x,y)+b.sub.l.sup.L)[T(x,y)-O.sub.l.sup.L (x,y)]

where;

N_p^l (x,y) is an input of the unit at said location (x,y);

W_p,q^l (x, y) is the connection weights;

O_p^l (x,y) is an output of the unit at said location (x,y);

l=(1,2, . . . L) is a layer number;

p=(1,2, . . . P^l) is a group number in a lth layer;

q=(1,2, . . . P^l +1) is a group number in a (l+1)th layer;

ƒ

(x) is a thresholding function;

ƒ

'"'"'(x) is a derivative of ƒ

(x);

Δ

w_p,q^l (x,y) is a change of the connection weight;

Δ

b_p^l is a change in a bias;

η

is a learning rate;

α

is a the constant factor of a moment term; and

T(x,y) is a desired output image.

15. A method as recited in claim 13, further comprising:

each unit in a group receiving input from a predetermined number of units in a preceding layer; and

selecting said predetermined number based upon a number of individual microcalcifications in a clustered microcalcification.

16. A method as recited in claim 13, further comprising:

processing said input data using a neural network having one input layer having one of said groups, a first hidden layer having two of said groups, a second hidden layer having one of said groups, and one output layer having one of said groups; and

each of said units receiving input from a predetermined number of units in a preceding layer.

17. A method as recited in claim 10, wherein said converting comprises:

correcting a background-trend of said regions of interest to produce corresponding background-trend corrected regions of interest;

normalizing said background-trend corrected regions of interest to produce corresponding normalized regions of interest;

scaling said normalized regions of interest into numerical data.

18. A method as recited in claim 17, wherein said correcting comprises:

fitting a predetermined curve to said regions of interest to produce corresponding estimated background trends; and

subtracting said estimated background trends from corresponding of said regions of interest;

and wherein said normalizing comprises;

##EQU8## where;

i_l (x,y) and i(x,y) denote said regions of interest after background-trend correction and after normalization, respectively; and

N² is a number of pixels in said regions of interest.

19. A method as recited in claim 10, further comprising:

said shift-invariant neural network outputting said output images having areas corresponding to detected microcalcifications;

counting said areas for each of said output images;

determining a region of interest to be a positive region of interest if a number of areas is at least a predetermined value.

20. A method as recited in claim 19, further comprising:

determining a region of interest to be a positive region of interest if said number of areas is at least two.

21. A method as recited in claim 10, wherein said detecting comprises:

image thresholding said output images to yield an output image at a predetermined threshold level of an output image intensity value; and

detecting microcalcifications using said image thresholded output images.

22. A method as recited in claim 10, further comprising:

performing image feature analysis on said regions of interest suspected of containing a clustered microcalcification.

23. A method as recited in claim 10, further comprising:

selecting second regions of interest in said digital mammogram corresponding to said clustered microcalcifications detected in said output images;

extracting image features from said second regions of interest;

performing image feature analysis on said extracted image features to identify clustered microcalcifications.

24. A method as recited in claim 23, wherein performing image feature analysis comprises feature thresholding.

25. A method as recited in claim 23, further comprising:

comparing clustered microcalcifications detected by both said shift-invariant neural network and said image feature analysis with true clustered microcalcifications; and

removing false-positive microcalcifications from said second regions of interest using said image feature analysis.

28. A method as recited in claim 10, wherein extracting said regions of interest comprises extracting spatial domain regions of interest.

26. A method as recited in claim 26, wherein removing said false-positives comprises:

feature thresholding said image feature analysis.

27. A method as recited in claim 26, wherein feature thresholding comprises performing at least one of:

determining a contrast of said detected microcalcification;

determining a size of said detected microcalcification;

determining a first moment of a power spectrum of said second regions of interest;

determining edge gradients in said second regions of interest;

determining cumulative edge gradients in said second regions of interest;

determining an average of an edge-gradient-orientation histogram of said second regions of interest;

determining a standard deviation of an edge-gradient-orientation histogram of said second regions of interest; and

determining an average optical density of said second regions of interest.

29. A system for detecting clustered microcalcifications in a digital mammogram, comprising:

a device for obtaining a digital mammogram;

a detector connected to said device which detects suspected microcalcifications in said digital mammogram;

a region of interest selector, connected to said device and said detector, which selects regions of interest in said digital mammogram corresponding to said suspected microcalcifications;

a region of interest processing device outputting processed regions of interest;

an input data scaler connected to said region of interest processing device for numerically scaling said processed regions of interest;

a shift-invariant neural network trained to detect clustered microcalcifications receiving as input data said numerically scaled regions of interest from said data scaler and outputting corresponding output images; and

a microcalcification detector for detecting microcalcifications in said output images.

30. A system as recited in claim 29, wherein said region of interest processing device comprises:

a background-trend corrector for correcting background-trends of said regions of interest; and

a pixel value normalizer for normalizing pixels of said background-trend corrected regions of interest.

31. A system as recited in claim 29, further comprising:

an image thresholder for image thresholding said output images.

32. A system as recited in claim 29, wherein said cluster detector comprises:

means for counting a number of clustered microcalcifications in said output image;

means for determining an output image to be positive if said number of clustered microcalcifications is at least a predetermined value.

33. A system as recited in claim 32, wherein said predetermined number is two.

34. A system as recited in claim 29, wherein said neural network comprises:

an input layer having a number of input units determined by a number of pixels in one half of power spectra of said regions of interest;

at least one hidden layer, wherein a number of hidden layers is determined using receiver operating characteristic analysis; and

an output layer.

35. A system as recited in claim 29, further comprising:

said neural network having an input layer having a plurality of input units, at least one hidden layer having a plurality of groups, each of said groups having a plurality of units, an output layer, and connection weights between said layers; and

connection weights between said layers being shift-invariant as

space="preserve" listing-type="equation">w(i,j;

x,y)=w(i-x, j-y)

where w(i,j;

x,y) denotes a connection weight between a unit at a location (i,j) in a first layer and a unit at a location (x,y) in a subsequent layer.

36. A system as recited in claim 29, further comprising:

means for training said neural network using a modified error back propagation method;

##EQU9## and in the case of said output layer,

space="preserve" listing-type="equation">δ

.sub.l.sup.L (x,y)=ƒ

'"'"'(N.sub.l.sup.L (x,y)+b.sub.l.sup.L) [T(x,y)-O.sub.l.sup.L (x,y)]

where;

N_p^l (x,y) is an input of the unit at said location (x,y);

W_p,q^l (x,y) is the connection weights;

O_p^l (x,y) is an output of the unit at said location (x,y);

l=(1,2, . . . L) is a layer number;

p=(1,2, . . . P^l) is a group number in a lth layer;

q=(1,2, . . . P^l +1) is a group number in a (l+1)th layer;

ƒ

(x) is a thresholding function;

ƒ

'"'"'(x) is a derivative of ƒ

(x);

Δ

w_p,q¹ (x,y) is a change of the connection weight;

Δ

b_p^l is a change in a bias;

η

is a learning rate;

α

is a the constant factor of a moment term; and

T(x,y) is a desired output image.

37. A system for detecting clustered microcalcifications in a digital mammogram, comprising:

a device for obtaining a digital mammogram;

a detector connected to said device which detects suspected microcalcifications in said digital mammogram;

a region of interest selector, connected to said device and said detector, which selects regions of interest in said digital mammogram corresponding to said suspected microcalcifications;

a region of interest processing device outputting a processed regions of interest;

an input data scaler connected to said region of interest processing device for numerically scaling said processed regions of interest;

a shift-invariant neural network trained to detect clustered microcalcifications receiving as input data said numerically scaled regions of interest from said data scaler and outputting corresponding output images;

a microcalcification detector for detecting microcalcifications in said output images; and

an apparatus for performing image feature analysis on microcalcifications detected by said microcalcification detector.

38. A system as recited in claim 37, wherein said region of interest processing device comprises:

a background-trend corrector for correcting background-trends of said regions of interest; and

a pixel value normalizer for normalizing pixels of said background-trend corrected regions of interest.

39. A system as recited in claim 37, further comprising:

an image thresholder for image thresholding said output images.

40. A system as recited in claim 37, wherein said cluster detector comprises:

means for counting a number of clustered microcalcifications in said output image;

means for determining an output image to be positive if said number of clustered microcalcifications is at least a predetermined value.

41. A system as recited in claim 40, wherein said predetermined number is two.

42. A system as recited in claim 37, wherein said neural network comprises:

an input layer having a number of input units determined by a number of pixels in one half of power spectra of said regions of interest;

at least one hidden layer, wherein a number of hidden layers is determined using receiver operating characteristic analysis; and

an output layer.

43. A system as recited in claim 37, further comprising:

said neural network having an input layer having a plurality of input units, at least one hidden layer having a plurality of groups, each of said groups having a plurality of units, an output layer, and connection weights between said layers; and

connection weights between said layers being shift-invariant as

space="preserve" listing-type="equation">w(i,j;

x,y)=w(i-x,j-y)

where w(i,j;

x,y) denotes a connection weight between a unit at a location (i,j) in a first layer and a unit at a location (x,y) in a subsequent layer.

44. A system as recited in claim 37, further comprising:

means for training said neural network using a modified error back propagation method;

##EQU10## and in the case of said output layer,

space="preserve" listing-type="equation">δ

.sub.l.sup.L (x,y)=ƒ

'"'"'(N.sub.l.sup.L (x,y)+b.sub.l.sup.L) [T(x,y)-O.sub.l.sup.L (x,y)]

where;

N_p^l (x,y) is an input of the unit at said location (x,y);

W_p,q^l (x,y) is the connection weights;

O_p^l (x,y) is an output of the unit at said location (x,y);

l=(1,2, . . . L) is a layer number;

p=(1,2, . . . P^l) is a group number in a lth layer;

q=(1,2, . . . P^l +1) is a group number in a (l+1)th layer;

ƒ

(x) is a thresholding function;

ƒ

'"'"'(x) is a derivative of ƒ

(x);

Δ

w_p,q^l (x,y) is a change of the connection weight;

Δ

b_p^l is a change in a bias;

η

is a learning rate;

α

is a the constant factor of a moment term; and

T(x,y) is a desired output image.

45. A system as recited in claim 37, wherein said apparatus for performing image feature analysis comprises:

a image feature extracter; and

an image feature classifier connected to said image feature extracter.

46. A system as recited in claim 45, wherein said image feature classifier comprises means for performing feature thresholding.

47. A system as recited in claim 37, further comprising:

a second region of interest selector connected to said microcalcification detector for selecting second regions of interest corresponding to said microcalcifications detected by said microcalcification detector;

a image feature extracter for extracting image features of said microcalcifications in said second regions of interest; and

an image feature classifier connected to said image feature extracter.

48. A system as recited in claim 47, further comprising:

a display for displaying said digital mammogram; and

wherein said image feature classifier further comprises means for indicating said detected microcalcifications on said displayed digital mammogram.