System for automated detection of cancerous masses in mammograms
First Claim
1. A method for automated analysis of a digitized mammogram to detect the presence of a possible cancerous mass comprising:
- detecting a region of interest (ROI) using said digitized mammogram and Fourier spatial bandpass analysis, said ROI corresponding with a possible cancerous mass, wherein a plurality of spatially bandpassed images of different resolutions corresponding with said digitized mammogram are employed to identify at least one brightness peak corresponding with the ROI;
extracting context data for said ROI from said digitized mammogram, said context data comprising attribute information determined in specific relation to the ROI; and
inputting image data corresponding with said ROI and said context data to a neural net trained in relation to the attributes of cancerous tissue regions, and generating an output from said neural net indicating whether a possible cancerous tissue mass is present in said ROI.
1 Assignment
0 Petitions
Accused Products
Abstract
A system for automated detection of cancerous masses in mammograms initially identifies regions of interest (ROIs) using Fourier analysis (e.g., by means of an optical correlator). Context data is extracted from the mammogram for each ROI, such as size, location, ranking, brightness, density, and relative isolation from other ROIs. The pixels in the ROI are averaged together to create a smaller array of super-pixels, which are input into a first neural net. A second neural net receives the output values from the first neural net and the context data as inputs and generates an output score indicating whether the ROI contains a cancerous mass. The second neural net can also be provided with context data from another view of the same breast, the same view of the other breast, or a previous mammogram for the same patient.
-
Citations
47 Claims
-
1. A method for automated analysis of a digitized mammogram to detect the presence of a possible cancerous mass comprising:
-
detecting a region of interest (ROI) using said digitized mammogram and Fourier spatial bandpass analysis, said ROI corresponding with a possible cancerous mass, wherein a plurality of spatially bandpassed images of different resolutions corresponding with said digitized mammogram are employed to identify at least one brightness peak corresponding with the ROI;
extracting context data for said ROI from said digitized mammogram, said context data comprising attribute information determined in specific relation to the ROI; and
inputting image data corresponding with said ROI and said context data to a neural net trained in relation to the attributes of cancerous tissue regions, and generating an output from said neural net indicating whether a possible cancerous tissue mass is present in said ROI. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25)
a first neural net receiving image data corresponding with said ROI as an input and generating at least one output value; and
a second neural net receiving said output values from said first neural net and said context data as inputs and generating an output indicating whether a possible cancerous mass is present in said ROI.
-
-
14. The method of claim 13, further comprising the substep of averaging pixels in said digitized mammogram corresponding with said ROI to obtain the image data, and wherein said input to said first neural net includes information corresponding with a size of said ROI.
-
15. The method of claim 14, wherein said context data comprises a least two types of information selected from a group consisting of:
- ROI location information, ROI size-related information, ROI brightness and density profile information, ROI isolation information relative to other detected regions of interest, ROI information relating to spicules therewithin.
-
16. The method of claim 15, wherein said context data is calculated from sine and cosine transformation of said at least two types of information.
-
17. The method of claim 1 wherein said neural net is initially trained using perturbations of a set of training images, said perturbations being generated by at least one of random scaling, rotation and translation of training images comprising said set of training images.
-
18. The method of claim 1, further comprising the step of reducing a resolution of said digitized mammogram for use in the detecting step.
-
19. The method of claim 1, further comprising the step of spatial bandpass filtering said digitized mammogram image to obtain said plurality of spatially bandpassed images, wherein a corresponding plurality of bandpass filters having differing resolutions are employed.
-
20. The method of claim 1, wherein said detecting step includes the substep of estimating, for at least one of said plurality of differing bandpassed images, the size of said at least one brightness peak based upon a determination of the number of pixels around said at least one brightness peak having a brightness within each of the following:
- a first predetermined percentage of a brightness of the peak in at least one bandpassed image;
a second predetermined percentage of a brightness of the digitized mammogram measured at the corresponding peak location; and
a third predetermined brightness percentage of a maximum brightness of the digitized mammogram.
- a first predetermined percentage of a brightness of the peak in at least one bandpassed image;
-
21. The method of claim 20, wherein a resolution of a said digitized mammogram is reduced.
-
22. The method of claim 20, wherein said estimating substep comprises the additional substep of calculating a radial size of said at least one brightness peak by measuring a distance from said peak in a plurality of directions until corresponding pixel brightness is one of the following:
- less than 1/exp(2) of the corresponding peak brightness;
less than a predetermined percentage of a brightness of the digitized mammogram measured at the corresponding peak location;
or greater than a predetermined percentage higher than the next pixel closer to the peak.
- less than 1/exp(2) of the corresponding peak brightness;
-
23. The method of claim 1, wherein said detecting step includes the substep of estimating, for at least one of the said plurality of differing bandpassed images, a size of said at least one brightness peak by comparing said at least one of said plurality of differing bandpassed images with said digitized mammogram.
-
24. The method of claim 1, further comprising the step of applying a predetermined plurality of spatial bandpass and spatial wavelet filters to obtain an enhanced image of said ROI.
-
25. The method of claim 24, wherein said enhanced image is employed to obtain at least a portion of said context data.
-
26. A system for automated detection of cancerous masses in mammograms comprising:
-
means for inputting a digital mammogram;
an optical processor for detecting a region of interest (ROI) using said digital mammogram and Fourier spatial bandpass analysis, said ROI corresponding with a possible cancerous mass, wherein a plurality of spatially bandpassed images of different resolutions corresponding with said digital mammogram are employed to identify at least one brightness peak corresponding with the ROI;
means for extracting context data for said ROI from said mammogram, said context data comprising attribute information determined in specific relation to the ROI; and
a neural net trained in relation to the attributes of cancerous tissue regions, for receiving image data corresponding with said ROI and said context data as inputs and for generating an output indicating whether said ROI contains a possible cancerous mass. - View Dependent Claims (27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40)
a first neural net receiving said ROI as an input and generating at least one output value; and
a second neural net receiving said output values from said first neural net and said context data as inputs and generating an output indicating whether said ROI contains a cancerous mass.
-
-
40. The system of claim 26 wherein neural net is initially trained using perturbations of a set of training images, said perturbations being generated by at least one of random scaling, rotation and translation of training images comprising said set of training images.
-
41. A system for automated detection of cancerous masses in mammograms comprising:
-
means for inputting a digital mammogram containing a first array of pixels;
an optical processor to detect a region of interest (ROI) using said digital mammogram and Fourier spatial bandpass analysis, said ROI corresponding with a possible cancerous mass, wherein a plurality of spatially bandpassed images of different resolutions corresponding with said digital mammogram are employed to identify at least one brightness peak corresponding with the ROI;
means for averaging pixels of said first array that correspond with only said ROI to create a second array of super-pixels;
a first neural net, trained in relation to the attributes of cancerous tissue regions, for receiving said second array of super-pixels as an input and generating at least one output value; and
a second neural net, trained in relation to the attributes of cancerous tissue regions, for receiving said at least one output value from said first neural net and at least a portion of said context data as inputs and generating an output indicating whether said ROI contains a possible cancerous mass. - View Dependent Claims (42, 43, 44, 45, 46, 47)
-
Specification