|
Software Patent Abstract
A method, system, and computer software product for correlating
medical images, comprising: obtaining first image data representative
of a first medical image including a first abnormality; obtaining
second image data representative of a second medical image including
a second abnormality; determining at least one feature value for
each of the first and second abnormalities using the first and second
image data; calculating, based on the determined feature values,
a likelihood value indicative of a likelihood that the first and
second abnormalities are a same abnormality; and outputting the
determined likelihood value.
Software Patent Claims
The invention claimed is:
1. A computer-implemented method for correlating medical images,
comprising: obtaining first image data representative of a first
medical image including a first abnormality; obtaining second image
data representative of a second medical image including a second
abnormality; determining at least one feature value for each of
the first and second abnormalities using the first and second image
data; calculating, based on the determined feature values, a likelihood
value indicative of a likelihood that the first and second abnormalities
are a same abnormality; and outputting the determined likelihood
value, wherein the step of obtaining the first image data comprises
obtaining first image data representative of the first medical image
derived using a first modality, the first modality being one of
mammography, sonography, and magnetic resonance imaging; and the
step of obtaining the second image data comprises obtaining second
image data representative of the second medical image derived using
a second modality different from the first modality.
2. The method of claim 1, wherein the step of obtaining the first
image data comprises obtaining first image data representative of
the first medical image using a first modality, in a given view;
and the step of obtaining the second image data comprises obtaining
second image data representative of the second medical image derived
using a second modality, in a view different from the given view.
3. The method of claim 1, wherein the step of obtaining the first
image data comprises obtaining first image data representative of
the first medical image using a first modality, in a given time;
and the step of obtaining the second image data comprises obtaining
second image data representative of the second medical image using
the first modality, at a time different from the given time.
4. The method of claim 1, wherein the step of obtaining the first
image data comprises obtaining first image data representative of
the first medical image using a first protocol of a first modality;
and the step of obtaining the second image data comprises obtaining
second image data representative of the second medical image using
a second protocol of the first modality, the second protocol being
different from the first protocol.
5. The method of claim 1, wherein the determining step comprises:
automatically segmenting the first and second abnormalities.
6. The method of claim 1, wherein the determining step comprises:
identifying at least one maximally correlated feature using a canonical
correlation analysis; and determining a value for each of the at
least one maximally correlated feature for each of the first and
second abnormalities.
7. The method of claim 6, wherein the identifying step comprises:
selecting the at least one maximally correlated feature from a list
of candidate features including filtered ARD, filtered margin sharpness,
posterior acoustic behavior, texture, NRG, average gray value, contrast,
and diameter.
8. The method of claim 1, wherein the determining step comprises:
evaluating a measure of similarity between maximally correlated
features to determine feature-conditioned likelihoods that the first
and second abnormalities are the same abnormality.
9. The method of claim 8, wherein the calculating step comprises:
using a likelihood ratio test based on the feature-conditioned likelihoods
that the first and second abnormalities are the same abnormality.
10. The method of claim 1, wherein the calculating step comprises:
applying the determined feature values to a classifier to obtain
the likelihood value.
11. A system for correlating medical images, comprising: a mechanism
configured to obtain first image data representative of a first
medical image including a first abnormality; a mechanism configured
to obtain second image data representative of a second medical image
including a second abnormality; a mechanism configured to determine
at least one feature value for each of the first and second abnormalities
using the first and second image data; a mechanism configured to
calculate, based on the determined feature values, a likelihood
value indicative of a likelihood that the first and second abnormalities
are a same abnormality; and a mechanism configured to output the
determined likelihood value, wherein the mechanism configured to
obtain the first image data comprises a mechanism configured to
obtain first image data representative of the first medical image
using a first modality, the first modality being one of mammography,
sonography, and magnetic resonance imaging; and the mechanism configured
to obtain the second image data comprises a mechanism configured
to obtain second image data representative of the second medical
image using a second modality different from the first modality.
12. The system of claim 11, wherein the mechanism configured to
obtain the first image data comprises a mechanism configured to
obtain first image data representative of the first medical image
using a first modality, in a given view; and the mechanism configured
to obtain the second image data comprises a mechanism configured
to obtain second image data representative of the second medical
image using the first modality, in a view different from the given
view.
13. The system of claim 11, wherein the mechanism configured to
obtain the first image data comprises a mechanism configured to
obtain first image data representative of the first medical image
using a first modality, at a given time; and the mechanism configured
to obtain the second image data comprises a mechanism configured
to obtain second image data representative of the second medical
image using the first modality, at a time different from the given
time.
14. The system of claim 11, wherein the mechanism configured to
obtain the first image data comprises a mechanism configured to
obtain first image data representative of the first medical image
using a first protocol of a first modality; and the mechanism configured
to obtain the second image data comprises a mechanism configured
to obtain second image data representative of the second medical
image using a second protocol of the first modality, the second
protocol being different from the first protocol.
15. The system of claim 11, wherein the mechanism configured to
determine comprises: a mechanism configured to automatically segment
the first and second abnormalities.
16. The system of claim 11, wherein the mechanism configured to
determine comprises: a mechanism configured to identify at least
one maximally correlated feature use a canonical correlation analysis;
and a mechanism configured to determine a value for each of the
at least one maximally correlated feature for each of the first
and second abnormalities.
17. The system of claim 16, wherein the mechanism configured to
identify the at least one maximally correlated feature comprises:
a mechanism configured to select the at least one maximally correlated
feature from a list of candidate features including filtered ARD,
filtered margin sharpness, posterior acoustic behavior, texture,
NRG, average gray value, contrast, and diameter.
18. The system of claim 11, wherein the mechanism configured to
determine comprises: a mechanism configured to evaluate a measure
of similarity between maximally correlated features to determine
feature-conditioned likelihoods that the first and second abnormalities
are the same abnormality.
19. The system of claim 18, wherein the mechanism configured to
calculate comprises: a mechanism configured to use a likelihood
ratio test based on the feature-conditioned likelihoods that the
first and second abnormalities are the same abnormality.
20. The system of claim 11, wherein the mechanism configured to
calculate comprises: a mechanism configured to apply the determined
feature values to a classifier to obtain the likelihood value.
21. A computer program product which stores, on a computer-readable
medium, instructions for execution on a computer system, which when
executed by the computer system, causes the computer system to perform
the steps of: obtaining first image data representative of a first
medical image including a first abnormality; obtaining second image
data representative of a second medical image including a second
abnormality; determining at least one feature value for each of
the first and second abnormalities using the first and second image
data; calculating, based on the determined feature values, a likelihood
value indicative of a likelihood that the first and second abnormalities
are a same abnormality; and outputting the determined likelihood
values, wherein the step of obtaining the first image data comprises
obtaining first image data representative of the first medical image
using a first modality, the first modality being one of mammography,
sonography, and magnetic resonance imaging; and the step of obtaining
the second image data comprises obtaining second image data representative
of the second medical image using a second modality different from
the first modality.
22. The computer program product of claim 21, wherein the step
of obtaining the first image data comprises obtaining first image
data representative of the first medical image using a first modality,
in a given view; and the step of obtaining the second image data
comprises obtaining second image data representative of the second
medical image using the first modality, in a view different from
the given view.
23. The computer program product of claim 21, wherein the step
of obtaining the first image data comprises obtaining first image
data representative of the first medical image using a first modality,
at a given time; and the step of obtaining the second image data
comprises obtaining second image data representative of the second
medical image using the first modality, at a time different from
the given time.
24. The computer program product of claim 21, wherein the step
of obtaining the first image data comprises obtaining first image
data representative of the first medical image using a first protocol
of a first modality; and the step of obtaining the second image
data comprises obtaining second image data representative of the
second medical image using a second protocol of the first modality,
the second protocol being different from the first protocol.
25. The computer program product of claim 21, wherein the determining
step comprises: automatically segmenting the first and second abnormalities.
26. The computer program product of claim 21, wherein the determining
step comprises: identifying at least one maximally correlated feature
using a canonical correlation analysis; and determining a value
for each of the at least one maximally correlated feature for each
of the first and second abnormalities.
27. The computer program product of claim 26, wherein the identifying
step comprises: selecting the at least one maximally correlated
feature from a list of candidate features including filtered ARD,
filtered margin sharpness, posterior acoustic behavior, texture,
NRG, average gray value, contrast, and diameter.
28. The computer program product of claim 21, wherein the determining
step comprises: evaluating a measure of similarity between maximally
correlated features to determine feature-conditioned likelihoods
that the first and second abnormalities are the same abnormality.
29. The computer program product of claim 28, wherein the calculating
step comprises: using a likelihood ratio test based on the feature-conditioned
likelihoods that the first and second abnormalities are the same
abnormality.
30. The computer program product of claim 21, wherein the calculating
step comprises: applying the determined feature values to a classifier
to obtain the likelihood value.
Mobile Phone Patent Description
BACKGROUND OF THE INVENTION
Field of the Invention
The present invention is directed to feature-based correlation
of lesions from multiple images. More precisely, the present invention
relates to the correlation of lesions observed on breast images
acquired from different modalities, which may include, e.g., mammography,
breast sonography, and magnetic resonance imaging (MRI). The present
invention also relates to the correlation of lesions observed on
breast images acquired from different views, times, or protocols,
of a single modality for a given patient.
The present invention also generally relates to computerized techniques
for automated analysis of digital images, for example, as disclosed
in one or more of U.S. Pat. Nos. 4,839,807; 4,841,555; 4,851,984;
4,875,165; 4,918,534; 5,072,384; 5,150,292; 5,224,177; 5,289,374;
5,319,549; 5,343,390; 5,359,513; 5,452,367; 5,463,548; 5,491,627;
5,537,485; 5,598,481; 5,622,171; 5,638,458; 5,657,362; 5,666,434;
5,673,332; 5,668,888; 5,732,697; 5,740,268; 5,790,690; 5,873,824;
5,881,124; 5,931,780; 5,974,165; 5,982,915; 5,984,870; 5,987,345;
6,011,862; 6,058,322; 6,067,373; 6,075,878; 6,078,680; 6,088,473;
6,112,112; 6,141,437; 6,185,320; 6,205,348; 6,240,201; 6,282,305;
6,282,307; 6,317,617 as well as U.S. patent application Ser. Nos.
08/173,935; 08/398,307 PCT Publication WO 96/27846); 08/536,149;
08/900,189; 09/027,468; 09/141,535; 09/471,088; 09/692,218; 09/716,335;
09/759,333; 09/760,854; 09/773,636; 09/816,217; 09/830,562; 09/818,831;
09/842,860; 09/860,574; 60/160,790; 60/176,304; 60/329,322; 09/990,311;
09/990,310; 09/990,377; 10/360,814; and 60/331,995; and PCT patent
applications PCT/US98/15165; PCT/US98/24933; PCT/US99/03287; PCT/US00/41299;
PCT/JS01/00680; PCT/US01/01478 and PCT/US01/01479, all of which
are incorporated herein by reference.
The present invention includes the use of various technologies
referenced and described in the documents identified in the following
LIST OF REFERENCES:
LIST OF REFERENCES
[1] E. A. Sickles, W. N. Weber, H. B. Galvin, S. H. Ominsky, and
R. A. Sollitto, "Baseline screening mammography: one vs two
views per breast," Am. J. Roentgenol., vol. 147(6), pp. 1149-53,
1986. [2] R. G. Blanks, S. M. Moss, and M. G. Wallis, "Use
of two view mammography compared with one view in the detection
of small invasive cancers: further results from the National Health
Service breast screening programme," J. Med. Screen., vol.
4(2), pp. 98-101, 1997. [3] N. Vujovic and D. Brzakovic, "Control
points in pairs of mammographic images," IEEE IP, vol. 6(10),
pp. 1388-1399, 1997. [4] Y. Kita, R. P. Highnam, and J. M. Brady,
"Correspondence between Different View Breast X-Rays Using
a Simulation of Breast Deformation," in Proceedings of the
IEEE Computer Society Conference on Computer Vision and Pattern
Recognition, June 23-25, pp. 700, 1998. [5] S. L. Kok-Wiles, J.
M. Brady, and R. P. Highnam, "Comparing mammogram pairs in
the detection of lesions," in Proceedings of the 4th International
Workshop on Digital Mammography, Nijmegen, The Netherlands, June
1998. [6] F. F. Yin, M. L. Giger, K. Doi, et al., "Computerized
detection of masses in digital mammograms: automated alignment of
breast images and its effect on bilateral-subtraction technique,"
Med. Phys. vol. 21(3), pp. 445-452, 1994. [7] N. Karssemeijer and
G. M. te Brake, "Combining single view features and asymmetry
for detection of mass lesions," in Proceedings of the 4th International
Workshop on Digital Mammography, Nijmegen, The Netherlands, June
1998, pp. 95-102. [8] R. Chandrasekhar and Y. Attikiouzel, "A
Simple Method for Automatically Locating the Nipple on Mammograms,"
IEEE Trans. Med. Imaging, vol. 16(5), pp. 483-494, 1997. [9] S.
van Engeland and N. Karssemeijer, "Matching Breast Lesions
in Multiple Mammographic Views," in Proceedings of MICCAI 2001:
4th International Conference, Utrecht, The Netherlands, October
14-17, pp. 1172, 2001. [10] F. Richard and L. Cohen, "A new
image registration technique with free boundary constraints: application
to mammography," Computer Vision and Image Understanding, vol.
89, pp. 166-196, 2003. [11] L. Hadjiiski, H. Chan, B. Sahiner, N.
Petrick, and M. A. Helvie, "Automated registration of breast
lesions in temporal pairs of mammograms for interval change analysis
local affine transformation for improved localization," Med.
Phys., vol. 28(6), pp. 1070-1079, 2001. [12] S. van Engeland, P.
Snoeren, J. Hendriks, and N. Karssemeijer, "A Comparison of
Methods for Mammogram Registration," IEEE Trans. Med. Imaging,
vol. 22(11), pp. 1436-1444, 2003. [13] G. Xiao, J. M. Brady, J.
A. Noble, M. Burcher, and R. English, "Non-rigid registration
of 3-D free-hand ultrasound images of the breast," IEEE Trans.
Med. Imaging, vol. 21, pp. 405-412, 2002. [14] J. F. Krucker, G.
L. LeCarpentier, J. B. Fowlkes, and P. L. Carson, "Rapid Elastic
Image Registration for 3D Ultrasound," IEEE Trans. Med. Imaging,
vol. 21, pp. 1384-1394, 2002. [15] Z. Huo, M. L. Giger, C. J. Vyborny,
U. Bick, P. Lu, D. E. Wolverton, and R. A. Schmidt, "Analysis
of spiculation in the computerized classification of mammographic
masses," Med. Phys., vol. 22, pp. 1569-1579, 1995. [16] M.
A. Kupinski and M. L. Giger, "Automated seeded lesion segmentation
on digital mammograms," IEEE Trans. Med. Imaging, vol. 17,
pp. 510-517, 1998. [17] M. L. Giger, H. Al-Hallaq, Z. Huo, C. Moran,
D. E. Wolverton, C. W. Chan, and W. Zhong, "Computerized analysis
of lesions in US images of the Breast," Acad. Radiol., vol.
6, pp. 665-674, 1999. [18] K. Horsch, M. L. Giger, L. A. Venta,
and C. J. Vyborny, "Automatic segmentation of breast lesions
on ultrasound," Med. Phys., vol. 28, pp. 1652-1659, 2001. [19]
K. Horsch, M. L. Giger, L. A. Venta, and C. J. Vyborny, "Computerized
diagnosis of breast lesions on ultrasound," Med. Phys., vol.
29, pp. 157-164, 2002. [20] Z. Huo, M. L. Giger, and C. J. Vyborny,
"Computerized analysis of multiple-mammographic views: potential
usefulness of special view mammograms in computer-aided diagnosis,"
IEEE Trans. Med. Imaging, vol. 20, pp. 1285-1292, 2001. [21] Z.
Huo, M. L. Giger, D. E. Wolverton, W. Zhong, S. Cumming, and O.
I. Olopade, "Computerized analysis of mammographic parenchymal
patterns for breast cancer risk assessment: feature selection,"
Med. Phys., vol. 27, pp. 4-12, 2000. [22] M. A. Kupinski, D. C.
Edwards, M. L. Giger, and C. E. Metz, "Ideal Observer Approximation
Using Bayesian Classification Neural Networks," IEEE Trans.
Med. Imaging, vol. 20, pp. 886-899, 2001.
[23] C. E. Metz and X. Pan, "Proper binormal ROC curves: theory
and maximum-likelihood estimation," J. Math. Psych., vol. 43,
pp. 1-33, 1999.
The entire contents of each reference listed in the LIST OF REFERENCES
are incorporated herein by reference.
DISCUSSION OF THE BACKGROUND
The merging of information from images obtained from mammography,
ultrasound, and MRI is an important problem in breast cancer diagnosis.
Breast images of a given patient can provide complementary information
regarding an abnormality. Moreover, comparing the images also improves
their interpretation.
Radiologists frequently compare breast images taken at different
times for identifying and evaluating lesions. Also, diagnostic mammography
and breast sonography are routinely performed to evaluate breast
abnormalities that are initially detected with screening mammography
or on physical examination.
However, comparing multiple images taken by the same modality,
but in different views or according to different configurations
of parameters, is a nontrivial task that requires considerable expertise
and experience on the part of radiologists. Indeed, there is considerable
difficulty in comparing multiple breast images or fusing information
from multiple breast images. This difficulty is due in part to the
images themselves, which provide non-identical, but complementary
information concerning the lesion. In addition, whereas more than
one lesion may be present in a patient in some cases, the totality
of the lesions may not be represented on all images obtained in
a study. Furthermore, lesions of the same abnormality can exhibit
different characteristics in different images due to differences
in view projection, imaging time, and/or contrast mechanism.
In light of these difficulties, computer-based techniques are needed
to correlate lesions observed on breast images acquired for a given
patient either using different modalities, such as mammography,
breast sonography, and MRI, or using a single modality, but different
views, times, or protocols for that modality. In fact, such techniques
are acutely needed since such correlations arise in numerous applications,
including applications using multi-view mammograms, sonograms, or
multi-modal images for breast cancer diagnosis. Such correlations
also arise in applications comparing breast images in longitudinal
studies for evaluation of disease diagnosis, prognosis, or patient
management for both human interpretation as well as CAD methods.
It is known that using at least two views of mammograms allows
radiologists to better detect and evaluate breast abnormalities
[1,2]. Studies have attempted to develop computer algorithms determining
geometrical transformations that can establish a one-to-one mapping
between two mammographic images.
Image registration techniques that have been considered for matching
mammogram pairs include identifying and matching control points
[3] and identifying and matching inherent image landmarks, such
as curves [4], regions [5], breast skin line [6,7] and nipple position
[8,9] to minimize energy functions defined by intensities and contours
of regions of interest [10].
Several image registration techniques have been applied to detect
lesions in temporal pairs of mammograms [11]. The purpose of such
registration methods [12], often used to compensate for differences
in breast compression, positioning and acquisition protocols on
mammograms, is to aid radiologist in detecting and analyzing changes
in the breast that may have occurred between the mammograms as well
as new asymmetries observed between the left and right breast.
Recent ultrasound-to-ultrasound breast image registration studies
[13,14] have concentrated on 3-D compound images. The volumes reconstructed
therein typically have deteriorated spatial resolution because of
the presence of refraction artifacts and tissue mis-registration
between sequential 2-D scans. Both studies use voxel-based intensity
correlation coefficients between two sub-volumes to implement a
deformation energy function and mutual information.
There is also a need for automated classification and computer-aided
diagnosis of mass lesions in breast. That is, one would like to
determine as reliably as possible whether breast lesions are benign
or malignant using computer-extracted features of the lesions found
on mammograms or sonograms [15,16,17,18].
Further, correlating lesions from two different breast mammogram
views of a given modality using CAD is generally accepted as a second
opinion to that of radiologists. It is highly desirable, when attempting
to differentiate between benign and malignant breast lesions using
a two-view mammogram or sonogram analysis, to have corresponding
lesions arranged in pairs to merge information. Due to the elasticity
of the breast tissue, deformation in the positioning and compression
of the examination procedure vary from one examination to the other.
As a result, finding the geometric relationship between breast images
poses a rather daunting task.
Consequently, deformable registration techniques typically require
a sophisticated model to obtain consistent deformations both locally
and globally. Iterative optimization techniques are often needed
to estimate geometric transformation parameters. However, the similarity
measurement or fitness function defined to that end is often highly
non-linear and contains many local maxima. Further, these optimization
techniques do not guarantee convergence to the global maximum. In
addition, these techniques are often computationally very expensive
and thus difficult to use in a clinical setting.
SUMMARY OF THE INVENTION
Accordingly, to overcome the problems of the related art, an embodiment
of the present invention provides a method, system, and medium for
performing feature-based correlation of lesions in multiple images.
According to an aspect of the present invention, there is provided
a method, system, and medium for correlating medical images, comprising:
(1) obtaining a first medical image including a first lesion; (2)
obtaining a second medical image including a second lesion; (3)
determining at least one feature value for each of the first and
second lesions using image data of the first and second medical
images; (4) calculating, based on the determined feature values,
a likelihood value indicative of a likelihood that the first and
second lesions are a same lesion; and (5) outputting the determined
likelihood value.
According to an aspect of the present invention, there is provided
a method, system, and medium for performing a feature-based correlation
of lesions obtained from different modalities including mammography,
sonography, and magnetic resonance imaging.
According to another aspect of the present invention, there is
provided a method, system, and medium for performing a feature-based
correlation of lesions obtained from different views, times, or
protocols for a single modality.
According to another aspect of the present invention, there is
provided a method, system, and medium for performing a correlation
of multiple lesions based on a single feature.
According to another aspect of the present invention, there is
provided a method, system, and medium for performing a correlation
of multiple lesions based on a plurality of features.
According to another aspect of the present invention, there is
provided a method, system, and medium for characterizing whether
lesions from a plurality of images correspond to the same lesion.
According to another aspect of the present invention, there is
provided a method, system, and medium for automatically characterizing
abnormalities using features that are traditionally used by radiologists
in the clinical evaluation of breast masses as well as lower-level
features that may not be as intuitive to the radiologist's eye-brain
system. These radiographic image features emerge within embodiments
of the present invention as mathematical descriptors of characteristics
of mass lesions and automatically correlate, i.e., match, lesions
observed on multiple breast images of the same patient.
Embodiments of the present invention provide new automated methods,
systems, and media employing an intelligent computer system/workstation
for computer-assisted interpretation of breast magnetic resonance
imaging medical images.
Embodiments of the present invention provide a classification scheme
for an automated matching of lesions from multiple breast images,
thereby aiding radiologists or oncologists, who are presented with
matching likelihood values in a human-readable form, to compare
and combine information from mammography, breast ultrasound, or
other image modalities to improve diagnostic accuracy and overall
patient outcome.
Embodiments of the present invention identify specific features
that achieve automatic lesion matching, thereby aiding radiologists
or oncologists to compare and combine information from multi-modality
breast images, and improve diagnostic accuracy and overall patient
outcome. Moreover, the features can be used without any information
regarding the location of a lesion characterized by those features.
Embodiments of the present invention also facilitate identification
of incorrectly matched lesions that belong to different patients.
Other methods, systems, and media of the present invention will
become apparent to one of ordinary skill in the art upon examination
of the following drawings and detailed description of the invention.
BRIEF DESCRIPTION OF THE DRAWINGS
A more complete appreciation of the invention and many of the attendant
advantages thereof will be readily obtained as the same becomes
better understood by reference to the following detailed description
when considered in connection with the accompanying drawings, wherein:
FIGS. 1A-1D illustrate four ultrasound images of two lesions in
the right breast of a patient;
FIGS. 2A and 2B illustrate two mammographic views, mediolateral
oblique (MLO) and craniocaudal (CC), of two lesions in the same
breast;
FIG. 3 illustrates a method for feature-based automated classification
and matching;
FIGS. 4A and 4B illustrate two-dimensional distributions of an
image feature generated from images taken in two different views
for corresponding and non-corresponding datasets;
FIG. 5 illustrates the distribution of likelihood for corresponding
and non-corresponding datasets for a posterior acoustic behavior
feature;
FIG. 6 illustrates a likelihood ratio histogram for corresponding
and non-corresponding datasets for a posterior acoustic behavior
feature;
FIG. 7 illustrates a ROC curve for round-robin analysis and a sonographic
feature (posterior acoustic behavior);
FIG. 8 illustrates a ROC curve for round-robin analysis and a mammographic
feature (average grey value);
FIG. 9 illustrates a canonical correlation analysis for multiple
features; and
FIG. 10 illustrates a system for correlating medical images.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
The aforementioned difficulties are mitigated by using image features
associated with the lesions and developing a classification scheme
for establishing lesion correspondence. That is, given two breast
images, lesions are first automatically segmented from the surrounding
breast tissues and a set of features (feature vector) is automatically
extracted from each identified lesion. Subsequently, for every two
indicated lesions that are not in the same image, a classifier will
examine their features and yield the likelihood that the two lesions
correspond to the same physical lesion.
Image features produced by automatic segmentation that have been
successful in developing computer-aided diagnosis (CAD) of breast
cancers are also useful for the task of lesion matching. An appropriate
subset of features useful for discriminating corresponding and non-corresponding
lesion pairs will need to be determined. The features will then
be used to develop discrimination methods for automated correlating
of lesions from multiple breast images.
Canonical correlation is a statistical analysis for identification
and quantification of associations between two sets of variables.
Canonical correlation works by finding linear combinations of variables
from each set, called canonical variables, such that the correlation
between the two canonical variables is maximized. The advantage
of this technique is to concentrate a high-dimensional relationship
between the two sets of variables into a few pairs of canonical
variables. According to an embodiment of the present invention,
canonical correlation analysis is performed to identify a subset
of features for the discrimination task.
To identify useful features for discrimination, one selects features
that have a higher canonical correlation coefficient generated by
the same physical lesion and have a lower correlation generated
by different lesions.
Database
Embodiments of the present invention will be discussed using certain
databases of medical images. These databases will now be described
to facilitate an understanding of the present invention. However,
it is to be understood that the use of these databases does not
limit the scope of the invention in any way and the correlation
analysis can be implemented for other types of medical images, such
as chest radiography, magnetic resonance imaging, etc.
First, consider a sonographic database of 262 biopsy-proven lesions.
Specifically, each sonographic biopsy-proven lesion includes both
transverse and longitudinal views, and the 262 lesions include 81
complicated cysts, 115 benign lesions, and 66 malignant lesions.
Second, consider a mammography database of 230 biopsy-proven lesions.
Specifically, each mammography biopsy-proven lesion includes craniocaudal
(CC) and mediolateral oblique (MLO) views, and the lesions include
112 complicated cysts, 8 benign solid lesions, and 110 malignant
lesions.
Further, in order to mimic a lesion mismatch scenario that might
occur in clinical practice, consider also three "non-corresponding"
datasets denoted by A, B, and C.
Dataset A contains lesion pairs obtained from patients who had
two or more lesions shown on different image views. One lesion from
one view was paired with the other lesion in another view, of the
same patient, as a non-corresponding lesion pair. The sonographic
database comprises 35 patients who had two or more lesions shown
on transverse and longitudinal views, which leads to 53 non-corresponding
lesion pairs. The mammographic database comprises 11 patients who
had two or more lesions shown on transverse and longitudinal views,
which leads to 25 non-corresponding lesion pairs.
FIGS. 1A-1D illustrate four ultrasound images. The images represent
two lesions, taken in both transverse and longitudinal views, from
the right breast of a patient. FIG. 1A shows a transverse image
for the lesion at location 10:00 B. FIG. 1B shows a longitudinal
image for the lesion at location 1:00 A. FIG. 1C shows a transverse
image for the lesion at location 1:00 A. FIG. 1D shows a longitudinal
image for the lesion at location 10:00 B. The same lesion, shown
in different views, has different shape and orientation. The arrow
indicates the correspondence between scans belonging to the same
physical lesion.
FIGS. 2A and 2B illustrate two mammographic lesions in the same
breast. FIG. 2A shows a MLO view and FIG. 2B shows a CC view. Again,
the same lesion, shown in different views, has a different shape
and orientation since the images are taken from different angles.
The arrow indicates the correspondence between scans belonging to
the same physical lesion.
There are few patients having two or more lesions in the same breast
in the non-corresponding database. The non-corresponding dataset
must therefore be extended by generating lesion pairs across patients.
To that end, dataset B contains lesion pairs obtained from different
patients who have lesions with the same pathology and dataset C
contains lesion pairs obtained from different patients who have
lesions with different pathology.
One approach to constitute these "non-matching lesion"
datasets is to generate all possible lesion combinations across
patients with the same lesion pathology, or with the different pathology,
and then randomly select 300 pairs among all possible pairs.
Table I summarizes the number of lesion pairs contained in each
dataset for both databases.
TABLE-US-00001 Corresponding Non-corresponding Datasets (pairs)
Database Dataset (pairs) A B C Sonography 262 53 300 300 Mammography
230 25 300 300
A Classification Framework for Lesion Matching
Conventional approaches for matching lesions generally begin with
an attempt to spatially align two images in order to establish a
one-to-one pixel correspondence between the images. Once the images
are aligned, establishing lesion correspondence can be relatively
straightforward. Unfortunately, an adequate spatial alignment between
two breast images is often very difficult to achieve due to the
non-rigid morphology of the breast and the differences in imaging
protocols. For example, breast compression is applied in obtaining
mammograms but not in breast ultrasound. Moreover, it is also rather
impractical to exactly reproduce the compression applied in a mammography.
Therefore, a breast can exhibit significantly different shapes on
two images. In addition, while an entire breast is typically imaged
in mammography, only a portion of the breast is usually scanned
in ultrasound, further complicating the task of spatially aligning
two such breast images. These difficulties associated with the spatial
alignment of breast images can be mitigated by using local image
features associated with the lesions for establishing lesion correspondence.
To achieve this, one can use image features produced by automatic
classification schemes. These image features have been successful
used in developing computer-aided diagnosis (CAD) of breast cancers
and are also suitable for the task of lesion matching.
Classification has been widely applied in CAD and image analysis.
Typically, once a suspicious region is detected, features characterizing
the region are extracted. These features, or a subset of them, are
then employed in an automatic classifier to yield an estimate of
the probability of malignancy.
FIG. 3 illustrates generally a method for lesion matching. Instead
of a geometric transformation establishing the correct one-to-one
mapping between (x, y) locations in two image, an embodiment the
method uses lesion characteristics automatically extracted from
lesions in the two images to distinguish between corresponding and
non-corresponding lesion pairs. In FIG. 3, two images are acquired
in steps 301a and 301b, respectively. In steps 302a and 302b, lesions
are extracted from the images by automatic segmentation from the
surrounding breast tissues. This can be done using any conventional
segmentation technique. Naturally, any kind of segmentation approach
could also be used within the scope of the present invention. In
steps 303a and 303b, feature vectors are extracted corresponding
to the automatically segmented lesions. Subsequently, in step 304,
a classifier evaluates the similarity between the features of the
lesions. In step 305, the likelihood that the two lesions correspond
to the same physical lesion is determined.
The lesion-matching step performed in step 304 will now be described
in more depth. Let {right arrow over (x)}.sub.1 denote the feature
vector extracted for a lesion seen in one image, and let {right
arrow over (x)}.sub.2 denote the feature vector extracted for a
lesion seen in another image. A classifier examines the "similarity"
between the feature vectors {right arrow over (x)}.sub.1 and {right
arrow over (x)}.sub.2 to yield an estimate of the likelihood that
the two lesions correspond to the same physical lesion seen on two
different images. An appropriate "similarity" measure
is learned from two datasets of the feature pairs, one derived from
the same physical lesions seen in different images, the "corresponding
dataset," and the other derived from physically different lesions
seen in different images, the "non-corresponding dataset."
In general, the two images can be generated either using the same
imaging modality or using different imaging modalities. In both
cases, the images can be taken in different views and/or using different
imaging parameters. In addition, the classification task can consider
multiple image features. Consider a single image feature for each
image, denoted x.sub.1 and x.sub.2, respectively, in designing the
classifier. Extensions to multi-element feature vectors and multi-modality
lesion matching naturally fall within the scope of the present invention.
The two-dimensional (2D) distributions of the feature pair (x.sub.1,
x.sub.2) show that the feature pair derived from corresponding datasets
exhibit a linear relationship. Therefore, in designing a classifier,
applying linear regression is appropriate. Subsequently, one obtains
the likelihoods for a given feature pair based on the distance from
the derived regression line for pairs corresponding to the same
physical lesions and for those from physically different lesions.
Classification of whether two lesions seen on two images correspond
to the same physical lesion is based, for example, on a likelihood
ratio test. The performance of the proposed classification scheme
is evaluated by receiver operating characteristics (ROC) analysis
[23].
Modeling with Linear Regression and Feature Selection
FIGS. 4A and 4B illustrate the two-dimensional distribution of
an image feature generated from sonographic images taken in the
transverse and longitudinal views for corresponding and non-corresponding
datasets. FIG. 4A relates to the sonographic image feature posterior
acoustic behavior extracted from lesions seen on the transverse
and longitudinal views. FIG. 4B relates to the mammographic image
feature average gray value extracted from lesions seen on the MLO
and CC views.
A linear relationship is observed for feature pairs extracted from
the corresponding dataset. In comparison, feature pairs derived
from the non-corresponding dataset show a much wider spread in distribution.
This difference between the spread of the distribution for the
two populations can be utilized for classification. To that end,
one can compute a correlation coefficient r for the 2D feature distribution
generated from the corresponding dataset and a correlation coefficient
r' for the distribution generated from the non-corresponding datasets.
The correlation coefficient may be, for example, Pearson's coefficient.
One can also obtain p-values of the derived correlation coefficients.
This procedure allows the identification of a number of candidate
features that may be useful for the proposed lesion-matching task.
For example, one may select features that yield r.gtoreq.0.5 with
p.ltoreq.0.05 while r'.ltoreq.0.5.
Principal component analysis (PCA) provides a useful means for
identifying the linear relationship observed in FIGS. 4A and 4B.
Given the joint distribution of n zero-mean random variables, z.sub.1,
. . . , z.sub.n, PCA identifies the orthonormal vectors .sub.1,
. . . , .sub.n (called the principal axes) so that the distributions
of {circumflex over (z)}.sub.i={right arrow over (z)} .sub.1, i=2,
. . . , n, where {right arrow over (z)}=(z.sub.1, . . . , z.sub.n),
are mutually uncorrelated. Furthermore, the distribution of {circumflex
over (z)}.sub.1 has the largest variance among all unitary transforms
of {right arrow over (z)}, i.e., the 1D subspace spanned by .sub.1
contains the most significant statistical variations of the distribution.
Similarly, for i=2, . . . , n, {circumflex over (z)}.sub.1 is the
unitary transform of {right arrow over (z)} that has the largest
variance while are statistically uncorrelated to {circumflex over
(z)}.sub.1, . . . , {circumflex over (z)}.sub.i-1; i.e., the 1D
subspace spanned by .sub.1 contains the most statistical variations
of the distribution in the subspace orthogonal to the subspace spanned
by .sub.1, . . . , .sub.i-1. Consider now the application of PCA
to the 2D distributions of the feature pair (x.sub.1, x.sub.2) derived
from the corresponding database.
Before applying PCA, the mean of the distribution ( x.sub.1, x.sub.2)
is calculated and subtracted from the data thereby yielding a zero-mean
population ({tilde over (x)}.sub.1, {tilde over (x)}.sub.2), where
{tilde over (x)}.sub.1=x.sub.1- x.sub.1 and {tilde over (x)}.sub.2=x.sub.2-
x.sub.2. The most significant statistical variations of the distribution
of ({tilde over (x)}.sub.1, {tilde over (x)}.sub.2) will therefore
appear in the first principal axis .sub.1. Conversely, the spread
of the distribution of the feature pair derived from the corresponding
dataset along the second principal axis .sub.2 is minimized among
all possible axes in the {tilde over (x)}.sub.1-{tilde over (x)}.sub.2
space. However, the spread of the distribution of the feature pair
derived from the non-corresponding dataset is not necessarily minimized
along this axis.
Therefore, a large difference in the spreads of the distributions
of the two populations is likely to be observed along .sub.2. Let
.sub.1=(e.sub.11, e.sub.12), e.sub.11.noteq.0. The regression line
in the x.sub.1-x.sub.2 space corresponding to the .sub.1 axis in
the {tilde over (x)}.sub.1-{tilde over (x)}.sub.2 space is given
by
.times..times..times..times..times..times..beta..beta..times..times..times-
..beta..times..times..times..times..times..beta. ##EQU00001## Likelihood
Ratio Test
The task of a two-group automated classifier is to map a multidimensional
observation variable into a scalar decision variable from which
a threshold can be derived to determine which group an observation
belongs to. A good candidate variable for discriminating feature
pairs generated from the corresponding dataset against those from
the non-corresponding dataset can be achieved by considering a function
of the distance of a feature pair to the regression line L. In particular,
one can consider the respective likelihood of a feature pair being
at a given distance from the regression line when the pair is derived
from the corresponding and the non-corresponding datasets. Discrimination
can then be based on a likelihood-ratio criterion [22].
Let H.sub.1 denote the hypothesis that a given feature pair (x.sub.1,
x.sub.2) is derived from the same physical lesion and H.sub.0 the
hypothesis that it is derived from physically different lesions.
Let d denote the distance of this feature pair to the regression
line established from the corresponding dataset, as described above.
The likelihood of H.sub.1 being true given the feature pair is the
conditional probability for obtaining the feature pair given the
hypothesis, i.e., l.sub.1(d)=prob(d|H.sub.1). Similarly, the likelihood
of H.sub.0 being true given the feature pair is l.sub.0(d)=prob(d|H.sub.0).
The likelihood ratio is given
.times..times..LAMBDA..function..function.d.function.d.function.d.function-
.d ##EQU00002##
This test statistic is compared against a pre-determined threshold
.OMEGA. for discrimination. The hypothesis H.sub.1 is declared true,
i.e., the given feature pair is declared to derive from the same
physical lesion, when .LAMBDA.(d).gtoreq..OMEGA.; otherwise, the
hypothesis H.sub.0 is declared true and the feature pair is considered
to derive from two physically different lesions.
To use this likelihood ratio test, the likelihood functions need
to be determined. Let (x.sub.1.sup.(i), x.sub.2.sup.(i)), i=1, .
. . , N, denote a feature pair in the corresponding dataset and
d.sup.(i), i=1, . . . , N, denote the distance of the pair to the
regression line L established from the corresponding dataset.
By definition, l.sub.1(d) is the distribution of d.sup.(i). Therefore,
it can be estimated from the histogram of d.sup.(i) provided that
the population size N is sufficiently large. Given a relatively
small population, one can apply gaussian smoothing to reduce sampling
errors. That is, one estimates l.sub.1(d) by
.function..times..pi..sigma..times..times..times..times..times..sigma.
##EQU00003## wherein .sigma.>0 is a parameter controlling a level
of smoothing. The likelihood function l.sub.0(d) is similarly estimated
by using feature pairs from the non-corresponding dataset. The Correlation
of Features Between Views
Fifteen features were automatically extracted to represent the
lesion in the sonographic database [19] to characterize the lesion's
shape, margin sharpness, echogenic texture, and posterior acoustic
behavior. A total of fifteen mammographic features [21] were also
extracted from each lesion to quantitatively characterize the lesion's
spiculation, margin sharpness, mass density, and texture. In order
to validate the assumption of a linear relation of features between
views, correlation of features were calculated between image views
for corresponding dataset and non-corresponding dataset A in both
databases. To that effect, Table II lists the top four correlation
coefficients r for the corresponding datasets and r' for the non-corresponding
datasets in the sonographic database. Similarly, Table II lists
the top four correlation coefficients r for the corresponding datasets
and r' for the non-corresponding datasets in the mammographic database.
The associated p-values for image features extracted from two views
are also listed in Tables II and III.
TABLE-US-00002 TABLE IV Corresponding Non-corresponding Datasets
Dataset A B C Feature r p-value r' p-value r' p-value r' p-value
Shape: Filtered 0.75 <0.00001 0.38 0.004 0.24 <0.001 -0.08
0.166 ARD Filtered margin 0.77 <0.00001 0.36 0.007 0.17 0.003
0.01 0.864 sharpness Posterior 0.82 <0.00001 0.14 0.311 0.24
<0.001 -0.10 0.095 acoustic behavior Texture: conY4 0.62 <0.00001
0.20 0.143 0.14 0.02 0.02 0.784
TABLE-US-00003 TABLE V Corresponding Non-corresponding Datasets
Dataset A B C p- p- p- p- Feature r value r' value r' value r' value
NRG.sub.ROI 0.47 <0.00001 0.31 0.144 0.15 0.011 -0.12 0.043 Average
0.66 <0.00001 0.07 0.731 0.07 0.233 0.01 0.917 gray value Contrast
0.56 <0.00001 0.29 0.165 0.03 0.608 0.07 0.225 Diameter 0.50
<0.00001 0.05 0.796 0.05 0.359 -0.10 0.088
Training the Classifier for Lesion Matching
After identifying the correlated feature vectors, one first uses
the features from the corresponding dataset to determine .beta..sub.0
and .beta..sub.1 for the linear regression model. The corresponding
and non-corresponding datasets are used to train the classifier.
FIGS. 5 and 6 illustrate the relationship between a feature obtained
from two different views for the three datasets in the mammographic
database and the sonographic database. FIG. 5 shows the likelihood
distribution of the corresponding and non-corresponding pairs in
terms of the sonographic feature posterior acoustic behavior. From
these two distributions, one can calculate the likelihood ratio
as a function of the distance to the regression line. Pairs of the
corresponding dataset tend to have a larger likelihood ratio, yielding
a likelihood ratio histogram skewed to the right, whereas pairs
of the non-corresponding dataset tend to have a smaller likelihood
ratio, yielding a likelihood ratio histogram skewed to the left.
One can then classify whether a pair of images corresponds to the
same actual lesion by performing the discrimination according to
this likelihood ratio. By varying the decision variable threshold
.OMEGA., an ROC curve can be computed from the training process
for the discrimination task. From the ROC curve, one can determine
an optimal threshold value of the classifier to yield some prescribed
sensitivity and specificity in the task of distinguishing between
corresponding and non-corresponding pairs.
The performance of this lesion-matching method can be evaluated
for individual computer-extracted features by calculating the area
under the curve (AUC) value of the ROC curve. ROC curves were generated
for both re-substitution and round-robin analysis. The round-robin
evaluation was performed to provide a more realistic estimate of
the performance. In a round-robin evaluation, one of the patients
is excluded from the database. That is, lesions from this patient
are removed from the database and classified according to a classifier
trained with the remaining lesions. This process is then repeated
for each patient. For datasets generated by lesions across patients,
one assumes lesions have been corresponding to a lesion and its
counter-part are from the same patient and eliminates them from
the training process.
Tests using the corresponding dataset paired with the non-corresponding
datasets A, B, and C were conducted to assess the performance of
the method in distinguishing between corresponding and non-corresponding
pairs. The results of the four features that yielded the four highest
AUC values are summarized in Tables IV and V which displays the
performance in terms of AUC of individual features for the sonographic
and mammographic databases, respectively, in the task of distinguishing
between corresponding and non-corresponding pairs. Independent validation
was performed 11 times by randomly selecting 300 non-corresponding
pairs for datasets B and C. The performance lists for these validations
are average AUC values resulting from ROC analysis of each of the
11 independent trials. The standard deviations on the AUC values
are given in parentheses. The posterior acoustic behavior outperformed
the other features in differentiating corresponding and non-corresponding
lesions identified on sonograms, yielding an AUC value ranging from
0.73 to 0.79 (0.72 to 0.78 in round robin). The average gray value
outperformed others for the mammographic database, yielding an AUC
value ranging from 0.68 to 0.71 (0.66 to 0.70 in round robin).
FIG. 7 illustrates the performance of the method in terms of ROC
curves for the posterior acoustic behavior feature for each test
in the task of distinguishing corresponding from non-corresponding
lesion pairs in the sonographic database. Similarly, FIG. 8 illustrates
the performance for the average gray value feature in the mammographic
database.
TABLE-US-00004 US Database: 262 Corresponding lesion pairs Data
set A Dataset B Dataset C Non-Corresponding Non-Corresponding Non-Corresponding
lesion pairs (53) lesion pairs (300) lesion pairs (300) Round Resubsituation
Round Resubsituation Round Features Resubsituation robin (.sigma.)
robin (.sigma.) (.sigma.) robin (.sigma.) Shape: Filtered 0.59 0.57
0.68 (0.01) 0.67 (0.01) 0.74 (0.01) 0.73 (0.01) ARD Filtered margin
0.61 0.58 0.71 (0.02) 0.70 (0.02) 0.76 (0.02) 0.74 (0.02) sharpness
Posterior 0.75 0.72 0.72 (0.01) 0.70 (0.02) 0.79 (0.02) 0.78 (0.02)
acoustic behavior Texture: conY4 0.67 0.6 0.69 (0.02) 0.67 (0.02)
0.70 (0.01) 0.68 (0.01)
TABLE-US-00005 Mammogram Database: 230 Corresponding lesion pairs
Dataset A Dataset B Dataset C Non-Corresponding Non-Corresponding
Non-Corresponding lesion pairs (25) lesion pairs (300) lesion pairs
(300) Round Resubsituation Round Resubsituation Round Features Resubsituation
robin (.sigma.) robin (.sigma.) (.sigma.) robin (.sigma.) NRG.sub.ROI
0.60 0.58 0.60 (0.01) 0.59 (0.01) 0.66 (0.01) 0.64 (0.01) Average
gray 0.68 0.66 0.70 (0.01) 0.69 (0.01) 0.72 (0.01) 0.71 (0.01) value
Contrast 0.68 0.53 0.62 (0.01) 0.60 (0.02) 0.63 (0.02) 0.61 (0.02)
Diameter 0.68 0.54 0.66 (0.02) 0.63 (0.02) 0.67 (0.01) 0.65 (0.02)
Furthermore, sonograms from 35 patients who have two or more lesions
in the same breast were used to conduct a test for 48 corresponding
lesion pairs and 53 non-corresponding lesion pairs. In this test,
the posterior acoustic behavior feature performed well with an AUC
of 0.81 (0.77 in round robin), which indicates that the non-corresponding
datasets B and C comprised by lesions across patients can provide
a good estimation in clinical practice.
FIG. 9 illustrates that the calculated canonical correlation coefficient
for four sonographic features (filtered ARD, filtered margin sharpness,
posterior acoustic behavior, and texture) derived from the two lesions
corresponding to the same physical lesion is 0.85. In comparison,
the calculated canonical correlation is 0.37 for the same features
when derived from different lesions. A linear relationship can be
observed for the canonical variables for features extracted from
the two lesions correspond to the same physical lesion. On the other
hand, the canonical variables for features derived from different
lesions show a much wider spread in distribution. The observed difference
in the spreads of the distributions for the two populations can
be utilized for classification.
FIG. 10 illustrates a system for carrying out embodiments of the
present invention. An imaging device 1001 is used to acquire medical
images. The images can be stored using a storage unit 1002. The
images can be processed by a computing unit 1003 comprising a lesion
segmentation device 1004, which automatically segments lesions from
the background, a feature extraction device 1005, which automatically
extracts at least one feature corresponding to the lesions, a similarity
evaluation device 1006, which determines the similarity between
the features corresponding to the lesions, and a likelihood estimation
device 1007 which determines a likelihood that the lesions all correspond
to the same lesion. The system can also comprise a computer-aided
diagnosis device 1008, a display device 1009, and/or a multimodality
device 1010, all configured to receive and use the likelihood that
the lesions all correspond to the same lesion.
Alternatively, the image data of interest may be stored in an image
archiving system, such as Picture Archiving Communications System
(PACS), and retrieved therefrom for processing according to the
present invention. Either way, the present invention obtains the
image data for subsequent processing as described before.
Embodiments of the invention can modify and improve upon a system
for a multimodality display workstation that displays the images
as well as information derived from the images. The novel workstation
can incorporate multimodality images of the same patient, automatically
assess whether the same lesion is being considered across the images,
and automatically relate the information calculated from one modality
to that from another. For example, the lesion seen on multiple images
may be characterized by mammographic feature, sonographic features,
MRI features, as well as combination features. Similar lesions across
modalities can be automatically retrieved from on-line multimodality
reference atlases. In addition, the physical lesion can be represented
on single and multi-modality distributions of malignant, benign,
and other states. Further, besides presentation of computer calculated
and/or generated data, novel means to display image data can also
be incorporated.
All embodiments of the present invention conveniently may be implemented
using a conventional general purpose computer or micro-processor
programmed according to the teachings of the present invention,
as will be apparent to those skilled in the computer art. Appropriate
software may readily be prepared by programmers of ordinary skill
based on the teachings of the present disclosure, as will be apparent
to those skilled in the software art. In particular, the computer
housing may house a motherboard that contains a CPU, memory (e.g.,
DRAM, ROM, EPROM, EEPROM, SRAM, SDRAM, and Flash RAM), and other
optional special purpose logic devices (e.g., ASICS) or configurable
logic devices (e.g., GAL and reprogrammable FPGA). The computer
also includes plural input devices, (e.g., keyboard and mouse),
and a display card for controlling a monitor. Additionally, the
computer may include a floppy disk drive; other removable media
devices (e.g. compact disc, tape, and removable magneto-optical
media); and a hard disk or other fixed high density media drives,
connected using an appropriate device bus (e.g., a SCSI bus, an
Enhanced IDE bus, or an Ultra DMA bus). The computer may also include
a compact disc reader, a compact disc reader/writer unit, or a compact
disc jukebox, which may be connected to the same device bus or to
another device bus.
Examples of computer program products associated with the present
invention include compact discs, hard disks, floppy disks, tape,
magneto-optical disks, PROMs (e.g., EPROM, EEPROM, Flash EPROM),
DRAM, SRAM, SDRAM, etc. Stored on any one or on a combination of
these computer readable media, the present invention includes software
for controlling both the hardware of the computer and for enabling
the computer to interact with a human user. Such software may include,
but is not limited to, device drivers, operating systems and user
applications, such as development tools. Computer program products
of the present invention include any computer readable medium which
stores computer program instructions (e.g., computer code devices)
which when executed by a computer causes the computer to perform
the method of the present invention. The computer code devices of
the present invention may be any interpretable or executable code
mechanism, including but not limited to, scripts, interpreters,
dynamic link libraries, Java classes, and complete executable programs.
Moreover, parts of the processing of the present invention may be
distributed (e.g., between (1) multiple CPUs or (2) at least one
CPU and at least one configurable logic device) for better performance,
reliability, and/or cost. For example, an outline or image may be
selected on a first computer and sent to a second computer for remote
diagnosis.
The invention may also be implemented by the preparation of application
specific integrated circuits or by interconnecting an appropriate
network of conventional component circuits, as will be readily apparent
to those skilled in the art.
The present method of feature-based correlation can also be implemented
more generally by one of ordinary skill in the art for the correlation
of other abnormalities of other organs. In particular, the present
method is applicable to any type of abnormalities in N dimensions
(N>1). Thus, an embodiment of the present method can be readily
applied to 2D/3D aneurysms, embolisms, lung cancer, stomach cancer,
etc.
For the purposes of this description, an image is defined to be
a representation of a physical scene, in which the image has been
generated by some imaging technology. Examples of imaging technology
include television or CCD cameras, or X-ray, sonar, nuclear, or
ultrasound imaging devices. The initial medium on which an image
is recorded could be an electronic solid-state device, a photographic
film, or some other device such as a photostimulable phosphor. That
recorded image could then be converted into digital form by a combination
of electronic (as in the case of a CCD signal) or mechanical/optical
means (as in the case of digitizing a photographic film or digitizing
the data from a photostimulable phosphor). The number of dimensions
that an image could have could be one (e.g., acoustic signals),
two (e.g., X-ray radiological images), or more (e.g., tomosynthesis
or nuclear magnetic resonance images).
Numerous modifications and variations of the present invention
are possible in light of the above teachings. It is therefore to
be understood that within the scope of the appended claims, the
invention may be practiced otherwise than as specifically described
herein.
|