Method for determining an optimally weighted wavelet transform based on supervised training for detection of microcalcifications in digital mammograms
First Claim
1. In a method for determining wavelet scale weighting coefficients, the improvement comprising:
- performing supervised training, including wavelet transformation, iterative inverse wavelet transformation and error analysis, on at least one training image derived from sampled radiation to determine three or more non-zero wavelet scale weighting coefficients, comprising,generating a teacher image based on a respective training image derived from sampled radiation, includingidentifying in said teacher image true locations of indicia in said respective training image, anddefining a training-free zone around each of the identified true locations of said indicia in said teacher image; and
selecting said three or more non-zero wavelet scale weighting coefficients based on iterative inverse wavelet transformation and error analysis performed in relation to each said training-free zone defined in said teacher image;
wherein said step of selecting said three or more non-zero wavelet scale weighting coefficients includes the substeps ofa) selecting first candidate weighting coefficients,b) generating a reconstructed training image based on the first candidate wavelet scale weighting coefficients and said respective training image,c) determining a first error using said reconstructed image and said respective teacher image,d) selecting second candidate weighting coefficients based on the first error and repeating steps b) and c) for the second candidate weighting coefficients until the error determined in step c) is within a specified error condition; and
wherein said substep of determining a first error includes the steps of;
determining a first error component based on a respective maximum value in each training-free zone in said at least one reconstructed image,determining a second error component based on differences between intensity values of said respective teaching image and said reconstructed training image outside the training-free zones,determining a ratio between a total number of said indicia identified and a number of pixels in the reconstructed training image outside the training-free zones,providing a balanced error by multiplying said ratio to the second error component, andcalculating a total error in said at least one reconstructed image by combining the balanced error and the first error component.
2 Assignments
0 Petitions
Accused Products
Abstract
A computer-aided diagnosis (CAD) method for detection of clustered microcalcifications in digital mammograms based on an image reconstruction using a substantially optimally weighted wavelet transform. Weights at individual scales of the wavelet transform are optimized based on a supervised learning method. In the learning method, an error function represents a difference between a desired output and a reconstructed image obtained from weighted wavelet coefficients of the wavelet transform for a given mammogram. The error function is then minimized by modifying the weights by means of a conjugate gradient algorithm. Performance of the optimally weighted wavelets was evaluated by means of receiver-operating characteristic (ROC) analysis which indicated that the present invention outperformed both a difference-image technique and partial reconstruction method currently used in CAD methods.
46 Citations
12 Claims
-
1. In a method for determining wavelet scale weighting coefficients, the improvement comprising:
-
performing supervised training, including wavelet transformation, iterative inverse wavelet transformation and error analysis, on at least one training image derived from sampled radiation to determine three or more non-zero wavelet scale weighting coefficients, comprising, generating a teacher image based on a respective training image derived from sampled radiation, including identifying in said teacher image true locations of indicia in said respective training image, and defining a training-free zone around each of the identified true locations of said indicia in said teacher image; and selecting said three or more non-zero wavelet scale weighting coefficients based on iterative inverse wavelet transformation and error analysis performed in relation to each said training-free zone defined in said teacher image; wherein said step of selecting said three or more non-zero wavelet scale weighting coefficients includes the substeps of a) selecting first candidate weighting coefficients, b) generating a reconstructed training image based on the first candidate wavelet scale weighting coefficients and said respective training image, c) determining a first error using said reconstructed image and said respective teacher image, d) selecting second candidate weighting coefficients based on the first error and repeating steps b) and c) for the second candidate weighting coefficients until the error determined in step c) is within a specified error condition; and wherein said substep of determining a first error includes the steps of; determining a first error component based on a respective maximum value in each training-free zone in said at least one reconstructed image, determining a second error component based on differences between intensity values of said respective teaching image and said reconstructed training image outside the training-free zones, determining a ratio between a total number of said indicia identified and a number of pixels in the reconstructed training image outside the training-free zones, providing a balanced error by multiplying said ratio to the second error component, and calculating a total error in said at least one reconstructed image by combining the balanced error and the first error component. - View Dependent Claims (2, 3, 4)
-
-
5. A computer readable medium on which is stored instructions that cause a computer system to determine wavelet scale weighting coefficients, by performing the step of:
-
performing supervised training, including wavelet transformation, iterative inverse wavelet transformation and error analysis, on at least one training image derived from sampled radiation to determine three or more non-zero wavelet scale weighting coefficients, comprising, generating a teacher image based on a respective training image derived from sampled radiation, including identifying in said teacher image true locations of indicia in said respective training image, defining a training-free zone around each of the identified true locations of said indicia in said teacher image, and selecting said three or more non-zero wavelet scale weighting coefficients based on iterative inverse wavelet transformation and error analysis performed in relation to each said training-free zone defined in said teacher image; wherein said step of selecting said three or more non-zero wavelet scale weighting coefficients includes the substeps of, a) selecting first candidate weighting coefficients, b) generating a reconstructed training image based on the first candidate wavelet scale weighting coefficients and said respective training image, c) determining a first error using said reconstructed image and said respective teacher image, d) selecting second candidate weighting coefficients based on the first error and repeating steps b) and c) for the second candidate weighting coefficients until the error determined in step c) is within a specified error condition; and wherein said substep of determining a first error includes the steps of; determining a first error component based on a respective maximum value in each training-free zone in said at least one reconstructed image, determining a second error component based on differences between intensity values of said respective teaching image and said reconstructed training image outside the training-free zones, determining a ratio between a total number of said indicia identified and a number of pixels in the reconstructed training image outside the training-free zones, providing a balanced error by multiplying said ratio to the second error component, and calculating a total error in said at least one reconstructed image by combining the balanced error and the first error component. - View Dependent Claims (6, 7, 8)
-
-
9. In an apparatus for determining wavelet scale weighting coefficients, the improvement comprising:
-
means for performing supervised training, including wavelet transformation, iterative inverse wavelet transformation and error analysis, on at least one training image derived from sampled radiation to determine three or more non-zero wavelet scale weighting coefficients, comprising, means for generating a teacher image based on a respective training image derived from sampled radiation, including means for identifying in said teacher image true locations of indicia in said respective training image, and means for defining a training-free zone around each of the identified true locations of said indicia in said teacher image; and means for selecting said three or more non-zero wavelet scale weighting coefficients based on iterative inverse wavelet transformation and error analysis performed in relation to each said training-free zone defined in said teacher image; wherein said means for selecting said three or more non-zero wavelet scale weighting coefficients comprises, a) means for selecting first candidate weighting coefficients, b) means for generating a reconstructed training image based on the first candidate wavelet scale weighting coefficients and said respective training image, c) means for determining a first error using said reconstructed image and said respective teacher image, and d) means for selecting second candidate weighting coefficients based on the first error and repeating use of said b) means for generating and said c) means for determining, for the second candidate weighting coefficients, until the error determined by said c) means for determining is within a specified error condition; and wherein said means for determining a first error comprises, means for determining a first error component based on a respective maximum value in each training-free zone in said at least one reconstructed image, means for determining a second error component based on differences between intensity values of said at least one teaching image and said at least one reconstructed training image outside the training-free zones, means for determining a ratio between a total number of said indicia identified and a number of pixels in the reconstructed training image outside the training-free zones, means for providing a balanced error by multiplying said ratio to the second error component, and means for calculating a total error in said at least one reconstructed image by combining the balanced error and the first error component. - View Dependent Claims (10, 11, 12)
-
Specification