Data processing: artificial intelligence – Neural network
Reexamination Certificate
1999-12-23
2003-04-22
Follansbee, John (Department: 2121)
Data processing: artificial intelligence
Neural network
C382S156000
Reexamination Certificate
active
06553356
ABSTRACT:
RESERVATION OF COPYRIGHT
This patent document contains material which is subject to copyright protection. The copyright owner has no objection to the facsimile reproduction by anyone of the patent document or the patent disclosure, as it appears in the U.S. Patent and Trademark Office patent files or records after the granting of a patent, but otherwise reserves all copyrights whatsoever.
BACKGROUND OF THE INVENTION
1. Field of Invention
This invention relates to computerized detection of abnormal anatomical regions depicted in radiographs.
2. Background
Detection and analysis of target objects in digital images are useful and important tasks. For example, detection and diagnosis of abnormal anatomical regions in radiographs, such as masses and microcalcifications in women's breast radiographs (mammograms), are among the most important and difficult tasks performed by radiologists.
Breast cancer is a leading cause of premature death in women over forty years old. Evidence shows that early detection, diagnosis and treatment of breast cancer significantly improves the chances of survival, reducing breast cancer morbidity and mortality. Many methods for early detection of breast cancer have been studied and tested, among them mammography. To date, mammography has proven to be the most cost effective means of providing useful information to diagnosticians regarding abnormal features in the breast and potential risks of developing breast cancer in large populations. The American Cancer Society currently recommends the use of periodic mammography and screening of asymptomatic women over the age of forty with annual examinations after the age of fifty. Mammograms may eventually constitute one of the highest volume X-ray images routinely interpreted by radiologists.
Between thirty and fifty percent of breast cancers detected radiographically demonstrate clustered microcalcifications on mammograms, and between sixty and eighty percent of breast cancers reveal microcalcifications upon microscopic examination. Therefore, any increase in the detection of clustered microcalcifications by mammography may lead to further improvements in its efficiency in the detection of early breast cancer.
Currently, acceptable standards of clinical care are that biopsies are typically performed on five to ten women for each cancer removed. With this high biopsy rate is the reasonable assurance that most mammographically detectable early cancers will be resected. However, reducing the biopsy rate without adversely affecting health is desirable. Accordingly, given the large amount of overlap between the characteristics of benign and malignant lesions which appear in mammograms, computer-aided detection and/or diagnosis (CAD) of abnormalities may have a great impact on clinical care.
At present, mammogram readings are performed visually by mammographic experts, that is, physicians and radiologists. Unfortunately, visual reading of mammograms has two major disadvantages. First, it is often possible to miss the breast cancer in its early stages. This is because, unlike many other cancers, there is as yet no clear way to detect premalignant changes in the breast. This results partly from the relative inaccessibility of breast tissue. A second disadvantage of visual reading of mammograms is that these readings are labor intensive, time consuming, and subjective. Also, multiple readings of a single mammogram may be necessary in order to increase the reliability of the diagnosis.
Therefore, it would be advantageous and useful to have CAD systems to help radiologists and physicians obtain quicker, more consistent, and more precise results when performing visual readings of mammograms. Such CAD systems would aid in cancer detection and improve the efficiency and accuracy of large-scale screening.
During the past twenty years, an ever-increasing number of CAD systems for mammography have been developed and tested in an attempt to increase diagnostic accuracy and improve the efficacy and efficiency of mammographic interpretations. Current CAD schemes for mass detection typically can be partitioned into three stages. The first stage identifies suspicious regions using either single image segmentation or bilateral image subtraction; the second stage calculates a feature vector for each of these suspicious regions; and the third stage classifies regions based on some type of decision mechanism applied to feature vectors. Regions which are ultimately classified as positive by the above process can be marked on a copy of the original image for presentation to a radiologist or for use in other analysis.
Recent reports on advances in CAD applied to mass detection indicate that there are some generalizations that can be made about performance limitations of current methods.
In order to achieve a high true-positive detection rate (i.e., sensitivity greater than 90%), all of these schemes report a relatively large false-positive detection rate, even when testing a limited image database. It is not uncommon to produce a false-positive detection rate of one region per-image for clustered microcalcifications and two false-positive detections per image for masses. Previous attempts to improve early CAD schemes have employed many different techniques, but none of these efforts have been able to reduce the false positive rate to acceptable levels.
It is widely believed that current performance of CAD for mass detection is significantly less than that of radiologists, given the same task, though quantitative comparisons are difficult because radiologists rarely read single images without supplementary information. To directly compare radiologists' performance with current CAD would require that radiologists be restricted to evaluating limited regions-of-interest on mammograms, in assessing the likelihood of a mass in the region.
SUMMARY OF THE INVENTION
In recent years, despite considerable effort by many groups, the rate of improvement in CAD performance has slowed to the point that performance statistics of the better systems seem to be approaching an asymptote. This performance level, which is largely independent of the specifics of implementation (e.g., neural networks, Bayesian networks or rule based systems), is believed to be well below the potential performance of CAD. The inventors of the present invention believe that a possible reason for this may be that essentially all current CAD implementations apply traditional paradigms of signal processing and pattern recognition to detect features in individual images, and it is probable that most of the relevant physical features, in single-views, have been identified and exploited to some extent. Whatever information remains untapped in single-views is either very elusive (i.e., difficult to program), at a higher level of abstraction, or has only a small potential impact on performance.
In contrast to CAD, mammographers routinely insist on concurrently reading multiple images (at least two of each breast) in evaluating cases. Apparently, a significant part of their decision process requires a synergistic interaction of multiple components of information, as opposed to evaluating each separately and then combining individual decisions. The degree to which each component of image information independently. affects the performance of radiologists, and potentially of CAD, is not known. A limited number of studies have shown that single-view mammography leads to a higher rate of recall, and results in a failure to detect 11% to 25% of cancers than would have been detected using multiple views. These observations, along with the fact that mammographers insist on comparing all views that are available, strongly suggests that they derive useful gains in performance from this.
Ipsilateral pairs of mammograms (i.e., two views of the same breast taken at some oblique angle) contain spatial information about a single breast. However, because the process of acquiring each image requires that the breast be compressed in a direction orthogonal to the image plane (flattened on the imag
Chang Yuan-Hsiang
Good Walter F
Gur David
Maitz Glenn S.
Wang Xiao Hui
Follansbee John
Holmes Michael B.
Pillsbury & Winthrop LLP
University of Pittsburgh - of the Commonwealth System of Higher
LandOfFree
Multi-view computer-assisted diagnosis does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Multi-view computer-assisted diagnosis, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Multi-view computer-assisted diagnosis will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3062326