Image analysis – Applications – Biomedical applications
Reexamination Certificate
1997-06-06
2001-06-12
Patel, Jay (Department: 2723)
Image analysis
Applications
Biomedical applications
C128S925000, C378S037000
Reexamination Certificate
active
06246782
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates generally to the field of computer-aided diagnosis of medical images. More specifically, the present invention discloses an automated system for detecting cancerous masses in digital mammograms.
2. Statement of the Problem
While the problem of reducing breast cancer mortality is substantial, mammography provides an important tool for early detection. False negative rates are high, however, resulting largely from the varying patterns of breast tissue and their ability to disguise a cancer. Systems that can rapidly scan and analyze many mammograms can help radiologists reduce such errors. However, to date, such systems have not been practical for clinical use because they have neither achieved sufficient performance (sensitivity and specificity) nor the required processing speed for analyzing mammograms in a cost-effective, near real-time manner.
Since no two breast lesions look the same, computer-aided diagnosis of mammograms must be capable of generalizing to make correct decisions on data patterns of lesions that the computer has never before experienced. This is analogous in many ways to locating camouflaged targets in that neither “target” is well defined, and detection may require looking for large scale features and distortions of the overall scene (or total breast).
The prior art in the field includes the following:
Doi et al. “Computer-aided Diagnosis: Development of Automated Schemes for Quantitative Analysis of Radiographic Images,”
Seminars in Ultrasound, CT, and MRI
1992, 13:140-152.
Nishikawa et al., “Performance of Automated CAD Schemes for the Detection and Classification of Clustered Microcalcifications,” In:
Digital Mammography
, Elsevier Science B. V., The Netherlands, 1994, 13-20.
Davies et al., “Automatic Computer Detection of Clustered Calcifications in Digital Mammograms,”
Phy. Med. Biol
. 1990, 35:8:1111-1118.
Ng et al., “Automated Detection and Classification of Breast Tumors,”
Computers and Biomedical Research
1992, 25:218-237.
Kegelmeyer, Jr. et al., “Dense Feature Maps for Detection of Calcifications,” In:
Digital Mammography
, Elsevier Science B. V., The Netherlands, 1994, 3-12.
Brettle et al., “Automatic Microcalcification Localisation using Matched Fourier Filtering,”
Digital Mammography
, Elsevier Science B. V., The Netherlands, 1994, pp. 21-30.
The prior art includes a variety of systems to detect both microcalcifications and mass lesions. For microcalcification or mass classification, the prior art techniques include either the use of features extracted by human observers or computer-extracted features. The latter features for microcalcifications include shape analysis such as compactness, Fourier descriptors of image boundaries, and average distance between calcifications applied to the extracted features. For mass classification methods, features used include spiculations or irregular masses that are identified by local radiating structures or analysis of gradient histograms generated by seed growing and local thresholding methods, morphologically-based features and image texture. These features are classified using either neural networks or binary decision trees. Common to all of these approaches is limited testing, with the computationally- intensive nature of the process implied as one reason for less than full, comprehensive testing.
Researchers associated with the University of Chicago (or Arch Development Corp.) have been major contributors in the development of computer-aided diagnosis for mammogram analysis (e.g., Doi et al.). This group has developed an automatic scheme for detecting clustered microcalcifications and is evaluating a method for classification of clusters. One approach involves using a linear filter to improve the signal-to-noise ratio of microcalcifications. Signal-extraction criteria are imposed to distinguish true lesions from artifacts. The computer then indicates locations that may contain clusters of microcalcifications on the film. These investigators were able to demonstrate that use of the program improved radiologists' accuracy in detecting clustered microcalcifications. In the detection computer-aided diagnosis scheme, 78 digitized mammograms (half with, half without clusters) were processed and achieved 87% sensitivity to detection of clusters.
Davies et al. have described a system for detection of clusters of calcifications using image analysis techniques. Their method was based on a local area thresholding process, and was also able to identify other breast structures relative to the clusters of calcification. The results of their feasibility trial were a 100% true positive classification (for clusters), with an 8% false positive rate and 0% false negatives.
A more complex automated model is described by Ng et al. for the detection and classification of stellate lesions and circumscribed lesions. The relative strength of this system is that both stellate and circumscribed lesions appear as circular, bright masses with fuzzy boundaries but they may be difficult to distinguish by an automated system. Their pilot trial showed a high detection rate with a low false positive rate.
Kegelmeyer et al. have reported promising results when applying algorithms using binary decision trees and a dense feature map approach in identifying spiculated lesions and calcifications. Kegelmeyer et al. have achieved 100% sensitivity and 82% specificity when merging edge information identifying spicules with local texture measures, thus eliminating false-positive detections.
Brettle et al. demonstrated a true positive success rate of 100% and false positive and false negative rates of 0% each when using matched Fourier filtering in the frequency domain of mammographic images for detecting micro-calcification clusters. This application only used 15 segments of images, seven containing microcalcifications and eight without microcalcifications.
Other efforts by the National Institutes of Health and the National Science Foundation include: (i) gray scale image processing for better presentation of the image to the radiologist; (ii) applications of solutions of large-scale constrained optimization problems; (iii) adapting eye dwell time as a cuer; and (iv) spatial filters and signal extraction techniques for detection of microcalcifications; noise smoothing, edge enhancement, and structured background correction methods for detection of the mass boundary and characteristics such as size, density, edge sharpness, calcifications, shape, lobulation, and spiculation extracted from the mass and used for classification.
All of the aforementioned reports indicate some level of success, though most use single scale techniques, not multiscale which is essential to a high confidence result. From the data presented, it is difficult to compare the various results since there is no common data set. It is also difficult to assess the complexity of each data set.
3. Solution to the Problem.
The present approach is based on an innovative concept to detect patterns in medical images (e.g., mammograms) using a Fourier transform optical correlator for image analysis followed by a series of neural networks hosted on a digital computer to analyze the results. An optical processor allows higher-order bandwidth, multi-resolution, multi-orientation approaches for feature extraction and enhancement that are not feasible in feal time on a digital computer. A hybrid optical/digital computer approach ensures sufficient processing power at a moderate cost to accommodate discriminating algorithms and yet analyze a mammogram in a matter of seconds.
SUMMARY OF THE INVENTION
This invention provides a system for automated detection of cancerous masses in mammograms. The mammogram is digitized and regions of interest (ROIs) are detected using Fourier analysis (e.g., by means of an optical correlator). The pixels in the ROI are averaged together to create a smaller array of super-pixels that are input into a first neural net. Context data is extracted from the mammogram for
Oliver, Jr. Davis ReGoalt
Shapiro Gary Lee
Lockheed Martin Corporation
Marsh & Fischmann & Breyfogle LLP
Patel Jay
LandOfFree
System for automated detection of cancerous masses in... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with System for automated detection of cancerous masses in..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and System for automated detection of cancerous masses in... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2481114