U.S. patent application number 12/033789 was filed with the patent office on 2008-09-25 for automatic time-of-flight selection for ultrasound tomography.
This patent application is currently assigned to LOS ALAMOS NATIONAL SECURITY. Invention is credited to Lianjie Huang, Cuiping Li.
Application Number | 20080229832 12/033789 |
Document ID | / |
Family ID | 39773385 |
Filed Date | 2008-09-25 |
United States Patent
Application |
20080229832 |
Kind Code |
A1 |
Huang; Lianjie ; et
al. |
September 25, 2008 |
AUTOMATIC TIME-OF-FLIGHT SELECTION FOR ULTRASOUND TOMOGRAPHY
Abstract
Ultrasound sound-speed tomography requires accurate picks of
time-of-flights (TOFs) of transmitted ultrasound signals, however,
manual picking on large datasets is time-consuming. An improved
automatic TOF picker is taught based on the Akaike Information
Criterion (AIC) and multi-model inference (model averaging), based
on the calculated AIC values, to improve the accuracy of TOF picks.
The automatic TOF picker of the present invention can accurately
pick TOFs in the presence of random noise with average absolute
amplitude of up to 80% of the maximum absolute synthetic signal
amplitude. The inventive method is applied to clinical ultrasound
breast data, and compared with manual picks and amplitude threshold
picking. Test results indicate that the inventive TOF picker is
much less sensitive to data signal-to-noise ratios (SNRs), and
performs more consistently for different datasets in relation to
manual picking. The technique provides noticeably improved image
reconstruction accuracy.
Inventors: |
Huang; Lianjie; (Los Alamos,
NM) ; Li; Cuiping; (Troy, MI) |
Correspondence
Address: |
JOHN P. O'BANION;O'BANION & RITCHEY LLP
400 CAPITOL MALL SUITE 1550
SACRAMENTO
CA
95814
US
|
Assignee: |
LOS ALAMOS NATIONAL
SECURITY
Los Alamos
NM
|
Family ID: |
39773385 |
Appl. No.: |
12/033789 |
Filed: |
February 19, 2008 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
60901903 |
Feb 16, 2007 |
|
|
|
Current U.S.
Class: |
73/620 ; 600/448;
73/597 |
Current CPC
Class: |
G01B 17/00 20130101;
A61B 8/13 20130101; A61B 8/5207 20130101 |
Class at
Publication: |
73/620 ; 73/597;
600/448 |
International
Class: |
G01N 29/00 20060101
G01N029/00; A61B 8/13 20060101 A61B008/13 |
Goverment Interests
STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT
[0002] This invention was made with Government support under
Contract No.
[0003] DE-AC52-06NA25396, awarded by the Department of Energy. The
Government has certain rights in this invention.
Claims
1. A method of selecting time-of-flight (TOF) for ultrasound
tomography waveforms generated by a given ultrasound tomography
transmitter-receiver device directed on a tissue sample,
comprising: receiving a plurality of ultrasound waveforms from an
ultrasound tomography transmitter-receiver device; determining
Akaike Information Criterion (AIC) values within a predetermined
time window; and selecting TOF for each said ultrasound waveform in
response to the application of wavelet transforms searching said
time window.
2. A method as recited in claim 1, wherein said AIC is determined
as a best-model in which the AIC value is minimized.
3. A method as recited in claim 1: wherein said AIC value is
determined in response to multi-model averaging in which a weighted
average of models is generated; and wherein said weights for each
model are assigned in response to the relative accuracy of each
candidate model within the multiple models being considered.
4. A method as recited in claim 1, wherein said predetermined time
window comprises a timing window determined in response to
transmitter-receiver geometry and the sound speed in water.
5. A method as recited in claim 1, further comprising filtering to
eliminate outliers in the TOF picks.
6. A method as recited in claim 5, wherein said filtering comprises
median filtering.
7. A method as recited in claim 5, wherein said median filter has a
length customized to the time differences between picked TOFs and
the corresponding calculated TOFs in water based on the ring array
geometry.
8. A method as recited in claim 7, further comprising replacing
filtered out values with median values.
9. A method as recited in claim 1, further comprising comparing
TOFs of reciprocal transmitter-receiver pairs and adjusting the
associated TOF picks if they exceed a threshold.
10. A method as recited in claim 9, wherein said threshold is
selectable by a user based on individual requirements and data
quality needs.
11. A method as recited in claim 9, wherein said adjusting of TOF
picks comprises replacing the TOF and its reciprocal TOF with an
average of both TOF values.
12. A method as recited in claim 1, wherein determining said AIC
value comprises comparing AIC values to a series of models which
are previously specified.
13. A method as recited in claim 1: wherein said AIC value is
determined from, AIC(k)=k log (var(S(1, k)))+(N-k-1) log
(var(S(k+1, N))), where S(1, k) and S(k+1, N) are the two segments
in the selected time window; and the variance function "var(.)" is
determined from, var ( S ( i , j ) ) = .sigma. j - i 2 = 1 j - i l
= i j ( S ( l , l ) - S _ ) 2 , i .ltoreq. j , i = 1 , , N and j =
1 , , N ( 2 ) ##EQU00007## where S is the mean value of S(i,j).
14. A method as recited in claim 1, further comprising generating
ultrasound tomograph imaging in response to said TOF
selections.
15. A method as recited in claim 1, wherein said ultrasound
tomograph comprises ultrasonic breast tomography.
16. A method as recited in claim 1, wherein said method is
configured to provide operator-independent, automatic,
determination of TOFs for a set of ultrasonic signals.
17. A method as recited in claim 1, wherein said method selects
TOFs without necessitating manual picking of TOF timing points in
each of said plurality of ultrasonic waveforms.
18. A method of selecting time-of-flight (TOF) for ultrasound
tomography waveforms generated by a given ultrasound tomography
transmitter-receiver device directed on a tissue sample,
comprising: receiving a plurality of ultrasound waveforms from an
ultrasound tomography transmitter-receiver device; determining a
predetermined time window using the sound speed of water for the
given transmitter-receiver device; determining Akaike Information
Criterion (AIC) values for the received data within said
predetermined time window; calculating a weighted average model for
the signal segment; selecting a TOF for each said ultrasound
waveform in response to the application of wavelet transforms
searching said time window; applying a median filter to the TOF
selections; and correcting each TOF associated with said plurality
of ultrasound waveforms in response to the difference between
reciprocals.
19. An apparatus for processing for ultrasound tomography
waveforms, comprising: means for receiving a plurality of
ultrasound waveforms from an ultrasound tomography
transmitter-receiver device directed through a tissue sample; a
computer processor and memory coupled to said means; programming
executable on said processor for, determining a predetermined time
window, determining Akaike Information Criterion (AIC) values
within said predetermined time window, and selecting TOF for each
said ultrasound waveform in response to the application of wavelet
transforms searching said time window.
20. A computer-readable media executable on a computer apparatus
configured for processing ultrasound tomography waveforms,
comprising: a computer readable media containing programming
executable on a computer processor configured for processing
ultrasound tomography waveforms in response to receiving a
plurality of ultrasound waveforms from an ultrasound tomography
transmitter-receiver device directed through a tissue sample; said
programming executable on said processor configured for,
determining a predetermined time window, determining Akaike
Information Criterion (AIC) values within said predetermined time
window, and selecting TOF for each said ultrasound waveform in
response to the application of wavelet transforms searching said
time window.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] This application claims priority from U.S. provisional
patent application Ser. No. 60/901,903 filed on Feb. 16, 2007,
incorporated herein by reference in its entirety.
INCORPORATION-BY-REFERENCE OF MATERIAL SUBMITTED ON A COMPACT
DISC
[0004] Not Applicable
NOTICE OF MATERIAL SUBJECT TO COPYRIGHT PROTECTION
[0005] A portion of the material in this patent document is subject
to copyright protection under the copyright laws of the United
States and of other countries. The owner of the copyright rights
has no objection to the facsimile reproduction by anyone of the
patent document or the patent disclosure, as it appears in the
United States Patent and Trademark Office publicly available file
or records, but otherwise reserves all copyright rights whatsoever.
The copyright owner does not hereby waive any of its rights to have
this patent document maintained in secrecy, including without
limitation its rights pursuant to 37 C.F.R. .sctn. 1.14.
BACKGROUND OF THE INVENTION
[0006] 1. Field of the Invention
[0007] This invention pertains generally to ultrasound imaging, and
more particularly to automatic time-of-flight selection for
ultrasonic signals.
[0008] 2. Description of Related Art
[0009] Ultrasonic imaging is used in a wide variety of medical and
clinical applications. Image formation in ultrasonography is
provided in response to analysis of the time-of-flight and the
angle of incidence of the reflected ultrasound signals. It will be
recognized that multi-path reflections often arise between the
target object and the transducer, such as in response to highly
reflective acoustic interfaces. These multi-path reflections
interfere with proper image formation. For example, a prolongation
of time-of-flight can lead to overestimation of the target object
depth within the body. In addition, changes to the angle of
incidence of the incoming sound signals cause aliasing in the
calculated target object position. Therefore, in many cases the
reflected sound waves are subject to both straight-line propagation
as well as multi-path reflections at the same time. In order to
overcome these problems ultrasonic imaging techniques have been
developed in which a skilled operator can select which
time-of-flight values result in generating the proper image.
[0010] Manual selection of time-of-flight (TOF) during ultrasonic
imaging is presently considered the best method of achieving
optimum image quality. However, this operator dependent process is
very time-consuming when processing a large amount of ultrasound
data, such as in the case of medical ultrasound tomography which
may involve many thousand signals to be resolved. Inaccurate
time-of-fight picks can result in noisy reconstructed sound-speed
images with erroneous information about tumors, leading to wrong
cancer detection and diagnosis.
[0011] Accordingly, a need exists for a system and method for
operator independent time-of-flight selection for ultrasound
imaging. These needs and others are met within the present
invention, which overcomes the deficiencies of previously developed
ultrasound imaging systems and methods.
BRIEF SUMMARY OF THE INVENTION
[0012] A method and system is described for operator-independent
selection of time-of-flight (TOF) ultrasound signals, such as for a
clinical imaging system using ultrasound sound-speed tomography.
The invention provides a robust and computationally efficient
solution which can be applied to replacing operator selected
time-of-flights in various ultrasound systems. The automatic TOF
"picker" is based on the Akaike Information Criterion (AIC). A
preferred embodiment of the invention method utilizes an approach
termed multi-model inference (model averaging), which is based on
calculating AIC values across range of weighted models toward
improving the accuracy of TOF picks. In one aspect a median filter
is afterward utilized to eliminate outliers in the TOF picks. In
another aspect of the invention, if the sensor system is
symmetrical, such as a ring, then the reciprocal nature of signals
is compared with TOFs being adjusted accordingly, such as averaging
the reciprocal signals which exceed certain boundary
conditions.
[0013] The method and apparatus provides an operator-independent,
computationally efficient, and robust picker, which can accurately
and reliably pick time-of-flights for clinical ultrasound signals,
even for those with low signal-to-noise ratios.
[0014] The invention is amenable to being embodied in a number of
ways, including but not limited to the following descriptions.
[0015] One embodiment of the invention can be generally described
as a method of selecting time-of-flight (TOF) for ultrasound
tomography waveforms generated by a given ultrasound tomography
transmitter-receiver device directed on a tissue sample,
comprising: (a) receiving a plurality of ultrasound waveforms from
an ultrasound tomography transmitter-receiver device; (b)
determining Akaike Information Criterion (AIC) values within a
predetermined time window; and (c) selecting TOF for each the
ultrasound waveform in response to the application of wavelet
transforms searching the time window.
[0016] In one aspect of the invention, the AIC value can be
determined in response to the best-model in which the AIC value is
minimized. However, preferably, the AIC value is determined in
response to multi-model averaging in which a weighted average of
models is generated; and in which the weights for each model are
assigned in response to the relative accuracy of each candidate
model within the multiple models being considered.
[0017] The predetermined time window comprises a timing window
which is preferably determined in response to transmitter-receiver
geometry and the sound speed in water.
[0018] The set of TOF picks is preferably filtered to eliminate
outliers in the TOF picks, for example by utilizing median
filtering, which in one mode is configured to have a length
customized to the time differences between picked TOFs and the
corresponding calculated TOFs in water based on the ring array
geometry. In one mode of the invention, the filtered out values are
replaced with median values.
[0019] In one aspect of the invention, TOFs of reciprocal
transmitter-receiver pairs are compared and the values of the
associated TOF picks are adjusted if they exceed a given threshold.
In one mode the threshold can be selected by a user based on
individual requirements and data quality needs. In at least one
implementation, adjusting of the TOF picks comprises replacing the
TOF and its reciprocal TOF with an average of both TOF values.
[0020] In at least one implementation the AIC value is determined
by comparing AIC values based on a series of models which have been
previously specified.
[0021] Implementations of the present apparatus and/or method can
be incorporated into various ultrasound systems, such as those
generating ultrasound tomograph imaging in response to the TOF
selections. By way of example, these systems may be configured for
performing ultrasonic breast tomography. The inventive system and
method is configured to provide operator-independent, automatic,
determination of TOFs for a set of ultrasonic signals. In the
method and system of the invention, TOFs are selected without
necessitating manual picking of timing in each of the plurality of
ultrasonic waveforms.
[0022] At least one implementation of the invention comprises a
method of selecting time-of-flight (TOF) for ultrasound tomography
waveforms generated by a given ultrasound tomography
transmitter-receiver device directed on a tissue sample,
comprising: (a) receiving a plurality of ultrasound waveforms from
an ultrasound tomography transmitter-receiver device; (b)
determining a predetermined time window using the sound speed of
water for the given transmitter-receiver device; (c) determining
Akaike Information Criterion (AIC) values for the received data
within the predetermined time window; (d) calculating a weighted
average model for the signal segment; (e) selecting a TOF for each
the ultrasound waveform in response to the application of wavelet
transforms searching the time window; (f) applying a median filter
to the TOF selections; and (g) correcting each TOF associated with
the plurality of ultrasound waveforms in response to the difference
between reciprocals.
[0023] At least one implementation of the invention is an apparatus
for processing ultrasound tomography waveforms, comprising: (a)
means for receiving a plurality of ultrasound waveforms from an
ultrasound tomography transmitter-receiver device directed through
a tissue sample; (b) a computer processor and memory coupled to the
receiving means; (c) programming executable on the processor for,
(c)(i) determining a predetermined time window, (c)(ii) determining
Akaike Information Criterion (AIC) values within the predetermined
time window, and (c)(iii) selecting TOF for each the ultrasound
waveform in response to the application of wavelet transforms
searching the time window.
[0024] At least one implementation of the invention is a
computer-readable media executable on a computer apparatus
configured for processing ultrasound tomography waveforms,
comprising: (a) a computer readable media containing programming
executable on a computer processor configured for processing
ultrasound tomography waveforms in response to receiving a
plurality of ultrasound waveforms from at least one ultrasound
tomography transmitter-receiver device directed through a tissue
sample, in which the programming executable on the processor
configured for, (a)(i) determining a predetermined time window,
(a)(ii) determining Akaike Information Criterion (AIC) values
within the predetermined time window, and (a)(iii) selecting TOF
for each the ultrasound waveform in response to the application of
wavelet transforms searching the time window.
[0025] The present invention provides a number of beneficial
aspects which can be implemented either separately or in any
desired combination without departing from the present
teachings.
[0026] An aspect of the invention is an operator-independent method
of selecting time-of-flight signals within an ultrasonic imaging
device.
[0027] Another aspect of the invention is a time-of-flight
selection method which utilizes multi-model inference in selecting
time-of-flight.
[0028] Another aspect of the invention is a time-of-flight
selection method which uses calculated AIC values within
multi-modal inference.
[0029] Another aspect of the invention is a time-of-flight
selection method which uses a median filter to eliminate outliers
in the TOF picks.
[0030] Another aspect of the invention is a time-of-flight
selection method which utilizes wavelet-AIC according to a weighted
average model instead of the `best model`.
[0031] Another aspect of the invention is a time-of-flight
selection method which operates by removing outliers of the TOF
picks using filtering, such as a median filter.
[0032] Another aspect of the invention is a time-of-flight
selection method which performs no signal preprocessing which can
introduce signal distortion effects due to filtering and wavelet
de-noising.
[0033] A still further aspect of the invention is a method that can
be implemented as hardware, software, or computer readable media,
for processing waveforms in response to ultrasonic tissue
imaging.
[0034] Further aspects of the invention will be brought out in the
following portions of the specification, wherein the detailed
description is for the purpose of fully disclosing preferred
embodiments of the invention without placing limitations
thereon.
BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWING(S)
[0035] The invention will be more fully understood by reference to
the following drawings which are for illustrative purposes
only:
[0036] FIG. 1 is a flow diagram of time-of-flight selection
according to an aspect of the present invention.
[0037] FIG. 2A is a graph of absolute TOF differences between the
manual picks and amplitude threshold picks.
[0038] FIG. 2B is a graph of absolute TOF differences between the
manual picks and AIC TOF picks according to an aspect of the
present invention.
[0039] FIG. 3A-3B are graphs of TOF picks within high and low
signal-to-noise ratio images, showing a comparison of manual picks,
amplitude threshold picks and AIC picks.
[0040] FIG. 4 is a graph of TOF picks according to an aspect of the
present invention, showing arrival timing for a plurality of
signals.
[0041] FIG. 5A-5C are tomographic images acquired by X-ray CT (FIG.
5A), and TOFs according to AIC method (FIG. 5B) and the amplitude
threshold method (FIG. 5C).
[0042] FIG. 6A-6B are tomographic images acquired using the
sound-speed reconstruction method according to the present
invention (FIG. 6A) in comparison with sound-speed reconstruction
using amplitude threshold picks (FIG. 6B).
[0043] FIG. 7A is a schematic image of a ring transducer array from
which ultrasonic tomographic data was gathered.
[0044] FIG. 7B is a schematic diagram of ultrasound pulse
interaction within a ring transducer as shown in FIG. 7A, showing
scattering of the ultrasound field from transmitter to
receiver.
[0045] FIG. 8 is a graph comparing picks based on best model ("O"),
manual pick ("X"), and weighted average ("*"), showing a magnified
section of the signal containing the TOF picks.
[0046] FIG. 9A-9B are images of travel time differences between TOF
picks according to an aspect of the present invention and the
corresponding calculated TOFs in water based on the ring array
geometry in FIG. 9A and post processed version of that data in FIG.
9B, showing data being median filtered and reciprocal pair
checked.
[0047] FIG. 9C-9D are images of sound-speed reconstructions for an
ultrasound breast dataset corresponding to the TOF data in FIG. 9A
and FIG. 9B respectively.
[0048] FIG. 10A-110C are graphs of distortions arising from
band-pass filtering and wavelet de-noising of raw ultrasound data
(FIG. 10A), second order zero-phase Butterworth band-pass filtered
ultrasound data (FIG. 10B), and wavelet de-noised ultrasound data
(FIG. 10C) according to aspects of the present invention.
[0049] FIG. 11A-11B are graphs of amplitude and AIC value,
respectively, prior to adding of random noise within a high SNR
synthetic ultrasound waveform.
[0050] FIG. 12A-12B are graphs of amplitude and AIC value,
respectively, which are like those of FIG. 11A-11B, to which random
noise has been introduced.
[0051] FIG. 13 is a graph of five waveform snapshots of manual
picking windows, showing solid triangles at the location of the
manual TOF picks.
[0052] FIG. 14A-14B are graphs of absolute TOF differences between
manual picks and amplitude threshold picks (FIG. 14A), and absolute
TOF differences between manual picks and wavelet-AIC picks
according to the present invention (FIG. 14B).
[0053] FIG. 15A-15B are graphs of TOF pick comparisons between
manual picks, amplitude threshold picks and wavelet-AIC picks
according to the present invention, showing use in a high SNR
waveform (FIG. 15A), and a low SNR waveform (FIG. 15B).
[0054] FIG. 16 is a graph of an overlay of TOF picks according to
an embodiment of the present invention, showing dots indicating the
location of the TOF picks.
[0055] FIG. 17A-17C are tomography images of a breast phantom
compared with an X-ray CT scan (FIG. 17A), ultrasound image with
TOFs picked by the improved AIC picker according to the present
invention (FIG. 17B), and ultrasound image using amplitude
threshold picks (FIG. 17C).
[0056] FIG. 18A-18B are tomography images of low SNR data using TOF
picks according to the present invention (FIG. 18A), and using
amplitude threshold picks (FIG. 18B).
[0057] FIG. 19A-19B are tomography images of low SNR data using TOF
picks according to the present invention (FIG. 19A), and using
amplitude threshold picks (FIG. 19B).
[0058] FIG. 20 is a block diagram of an ultrasonic tomography
device according to an aspect of the present invention, shown for
processing a plurality of waveforms from which TOFs are
automatically selected.
DETAILED DESCRIPTION OF THE INVENTION
[0059] Referring more specifically to the drawings, for
illustrative purposes the present invention is embodied in the
apparatus generally shown in FIG. 1, 2B, 3A-3B, 4, 5B, 6B, 8-12B,
14B, 15A-15B, 16, 17B, 18A, 19A and 20. It will be appreciated that
the apparatus may vary as to configuration and as to details of the
parts, and that the method may vary as to the specific steps and
sequence, without departing from the basic concepts as disclosed
herein.
Section A
[0060] 1. Introduction.
[0061] The wavelet-AIC (Akaike Information Criterion) picker is
based on an autoregressive (AR) AIC picker that assumes an
ultrasound signal can be divided into locally stationary segments
and that the segments before and after the TOF represent two
different stationary processes. Classically, the wavelet-AIC picker
applies a running window and a wavelet transform to continuously
search for an appropriate time window for the final TOF picking.
Data points within the selected time window are divided into two
segments at each data point i (i=1, . . . , k, . . . , N, where N
is the total number of data points in the selected time window). To
calculate the AIC function directly from the waveform for a given
data point k the wavelet-AIC picker uses the formula of N. Maeda,
found in the publication "A Method for Reading and Checking Phase
Times in Autoprocessing System of Seismic Wave Data, Zisin",
Journal of Seismological Society of Japan 38 (1985) 365-379, to
yield the following:
AIC(k)=k log (var(S(1, k)))+(N-k-1) log (var(S(k+1, N))) (1)
where S(1,k) (for data points 1 through k) and S(k+1, N) (for data
points k+1 through N) are the two segments in the selected time
window, and the variance function "var(.)" is calculated using:
var ( S ( i , j ) ) = .sigma. j - i 2 = 1 j - i l = i j ( S ( l , l
) - S _ ) 2 , i .ltoreq. j , i = 1 , , N and j = 1 , , N ( 2 )
##EQU00001##
where S is the mean value of S(i,j). The AIC value given by Eq. (1)
measures the information loss of the selected model to approximate
reality.
[0062] The point with minimum AIC value (e.g., minimum information
loss, therefore referred to as `best model`) is selected to be the
TOF point in the wavelet-AIC auto-picker of H. Zhang, C. Thurber
and C. Rowe, in the article entitled "Automatic P-Wave Arrival
Detection and Picking with Multiscale Wavelet Analysis for
Single-Component Recordings" published in the Bulletin of the
Seismological Society of America, vol. 93 (2003) pages
1904-1912.
[0063] The automatic wavelet-AIC TOF picker concept is then further
improved according to the present invention by: (1) determining
TOFs according to a weighted average model instead of the `best
model`; (2) removing outliers of the TOF picks using a median
filter; and (3) eliminating effects of signal distortion due to
filtering and wavelet de-noising which arises during data
preprocessing, wherein signal-to-noise ratio (SNR) is improved.
[0064] 2. Using a Weighted Average Model to Determine TOFs.
[0065] FIG. 1 illustrates an embodiment 10 for an AIC picker method
of the invention. A predefined window is calculated (determined)
using the sound speed of water as represented by block 12. In block
14 the AIC value is calculated (determined) for the original data
within the predefined window. A weighted average is then computed
as per block 16 for the signal segment in the time window. From the
above information a time-of-flight is then determined at block 18,
and the process is repeated at block 20 for all waveforms. In block
22 a median filter is preferably applied to the TOFs, for example
with the filtered out (rejected) TOFs 24 being preferably replaced
at block 26 with median values. Accepted TOFs 28 and replaced
median values 26 are received in block 30 in which TOFs of
reciprocal transmitter-receiver pairs are compared. If the
difference values are less than or equal to a given threshold as
per block 32, then the TOFs are output at block 34. Otherwise, when
the differences are greater than the threshold as determined in
block 36, then the TOF values are discarded at block 38. The
following discusses in more detail aspects of the invention.
[0066] AIC value itself has no physical meaning and it becomes
valuable only when compared across a series of models which are a
priori specified. The model with the minimum AIC value is the best
among all models being compared ("best model"). The measure
associated with the AIC value that can be used to compare models is
the normalized Akaike weights, as found in Eqs. (3) and (4), which
indicate the relative importance of candidate models. In most
cases, the best model (corresponding to the minimum AIC value) may
have competitors for the top rank. An elegant solution to making an
inference based on the entire set of models is to compute the
weighted average based on the model uncertainties (i.e., Akaike
weights), which is termed model averaging or model inference.
[0067] The running window for the prior uses of a wavelet-AIC
picker are not necessary for use in clinical ultrasound data. An
appropriate time window used for the TOF pick can be well defined
based on the transmitter-receiver geometry and the sound-speed of
water since sound-speed of tissue (e.g., breast tissue) is very
close to that of water. The original geologic wavelet-AIC picker
picks the point corresponding to the best model within the
predetermined time window as the TOF. To incorporate all the
information near the best model, weighted model averaging is
applied to pick the TOF in the following sequence:
[0068] (1) Calculate AIC values (AIC.sub.i, i=1, . . . n) for a
series of data points near the point with the minimum AIC
(AIC.sub.min) using Eq. (1).
[0069] (2) Obtain the differences between AIC.sub.i(i=1, . . . n)
and AIC.sub.min:
.DELTA..sub.i=AIC.sub.i-AIC.sub.min (3)
[0070] (3) Compute the Akaike weights for each data point within
the time window:
w i = exp ( - .DELTA. i / 2 ) r = 1 n exp ( - .DELTA. r / 2 ) ( 4 )
##EQU00002##
[0071] (4) Obtain the TOF value using the weighted average:
t onset = i = 1 n w i t i ( 5 ) ##EQU00003##
where t.sub.i (i=1, . . . n) are the corresponding travel times for
those points discussed in (1)-(3) and wi is obtained using Eq.
(4).
[0072] When there is a sharp global minimum, the AIC value that
indicates a high SNR, the difference between the average model pick
and the best model pick is negligible. However, if the global
minimum is not very sharp, which indicates a low SNR, the weighted
average model can pick the TOF more accurately that the best model
pick. This is the primary advantage of the automatic TOF picking
method based on the weighted model averaging scheme.
[0073] 3. Removing Outliers of TOF Picks.
[0074] To remove outliers in the TOF picks, the TOF picks are first
compared for the reciprocal transmitter-receiver pair. The
reciprocal transmitter-receiver pair means that two transducers in
the ring array transmit and receive signals in opposite directions;
which is typically possible in that most sensor heads are
symmetrical, such as radially symmetrical. Ideally, the TOF picks
for a reciprocal pair should be the same, however, this rarely
occurs in practice. In the data cleaning process according to the
present invention, if the TOF difference for the reciprocal pair
exceeds a predefined threshold, then the TOF picks are adjusted,
for example replacing both TOF picks by the average of the two TOF
picks. The predefined threshold value can be customized by users
based on their individual requirements and data quality needs.
[0075] To eliminate outliers in the TOF picks more effectively,
this embodiment applies a median filter with a customized length to
the time differences (TD) between the TOFs picked according to the
invention and the corresponding calculated TOFs in water based on
the ring array geometry.
[0076] Although other techniques can be utilized without departing
from the teachings of the present invention, the median filter is a
particularly well-suited tool for reducing "salt and pepped" noise
(outliers). To take advantage of this property and the continuity
property of a TOF surface (formed by TOFs of all transmission
data), TDs are rearranged into a 2-D matrix in such a way that each
row represents the TD values for a single transmitter, and TD
values for adjacent transmitters in the ring array are put into
adjacent rows (except the first and last transmitter due to the
circular geometry of the ring). This rearrangement results in a 256
by 256 matrix (D). Another 256 by 256 matrix (M) containing all
median values is calculated with a sliding window of the same size
as the median filter. Adaptive thresholds for the median filter are
set up by calculating the standard deviation (STD) and the mean
value (ME) of TDs:
TolMin=ME-f*STD,
TolMax=ME+f*STD. (6)
where To/Min and To/Max are the minimum and maximum tolerance of
time differences, respectively, and f is a customized scale factor
of the standard deviations with a value between 0 and 1. The median
filter based on the above thresholds is applied to the matrix D: if
D(i,j)-M(i,j)<TolMin or D(i,j)-M(i,j)>TolMax, the
corresponding picked TOF is replaced with the median value;
otherwise, it is discarded.
[0077] 4. Eliminating Distortion of Filtering and Wavelet
De-noisin.
[0078] Filtering and techniques for elimination of noise
(de-noising techniques) can be utilized to preprocess a signal to
improve its signal-to-noise ratios (SNR). Original wavelet-AIC
pickers were characterized by applying the wavelet de-noising to
raw seismogram data before picking the TOF.
[0079] However, both the filtering and wavelet de-noising may
distort a signal while attempting to increase the SNR. Due to the
short transient time of ultrasound signals, a small distortion of
the ultrasound waveform may result in large unwanted artifacts or
erroneous information during TOF picking. For these reasons and in
order to preserve the true shapes of ultrasound signal onsets as
much as possible, a preferred embodiment of the present approach
does not perform any data preprocessing toward improving
signal-to-noise ratios (SNRs), because the improved automatic TOF
picker described herein is configured to handle ultrasound data
with low SNRs as demonstrated in the following.
[0080] 5. Capability to Handle Noisy Data.
[0081] To estimate the maximum amount of random noise that our
improved wavelet-AIC TOF picker can tolerate, progressively
increasing amounts of random noise were added to a synthetic
ultrasound waveform with a similar spectrum to that of in vivo
ultrasound breast data acquired by a ring transducer array. The
average amplitudes of the random noise were, respectively, 0%, 20%,
40%, 60% and 80% of the maximum absolute amplitude of the synthetic
ultrasound signal. The TOF picker according to the present
embodiment was able to consistently detect the correct TOF in the
presence of even 80% white noise, which corresponds to a 4.5 dB
SNR. This is not surprising as the testing performed herein, as
well as prior work on the AIC approach itself, indicate that AIC
TOF pickers should be highly tolerant of relatively high noise
levels if an appropriate time window is utilized.
[0082] 6. Experimental Results.
[0083] To assess the performance of the inventive improved AIC
picker, TOFs picks according to the invention were compared with
those of the traditional amplitude threshold picker, and manual
picks on 1160 waveforms of in vivo ultrasound breast data acquired
using a ring transducer array. Manual picking was conducted by an
experienced expert who determined the TOF of each waveform by
recognizing the first rise time of the signal, and thus serves as a
standard of comparison.
[0084] FIG. 2A-2B illustrate comparisons of TOF picking. FIG. 2A
illustrates the absolute time difference between manual picks and
amplitude threshold picks. FIG. 2B illustrates absolute time
difference between manual picks and wavelet-AIC TOF picks according
to the present invention. The statistics illustrate that for these
1160 in vivo ultrasound breast waveforms, over 85% of the TOFs
picked by the improved wavelet-AIC picker of the present invention
are within the three sample points from the manual picks. Among
picks by the amplitude threshold picker, only 48% of them are
within the three sample points from the manual picks. The dashed
lines indicate the three sample point interval from manual
picks.
[0085] Further analysis with more ultrasound data from breast
examinations reveal that the performances of the improved
wavelet-AIC picker of the current invention and amplitude threshold
picker are more comparable for less noisy data, while in response
to high noise level (or low SNRs) data, the accuracy of the
amplitude threshold picker drops abruptly.
[0086] FIG. 3A-3B illustrates two waveforms of different SNRs along
with TOF picks utilizing picking by manual methods, amplitude
threshold, and AIC techniques according to the present invention.
The manual picks are designated by an "X", the weighted average
picks with an "O", and the AIC picks marked by a "*". It will be
noted that the AIC picks remain close to that of the manually
selected standard of comparison. It can also be clearly seen that
for the high SNR waveform of FIG. 3A, the three TOF picks are
reasonably close to one another, although the improved AIC pick of
the invention are closer to the manual pick. FIG. 3B illustrates a
low SNR waveform showing a noisier waveform in which the amplitude
threshold picker selected the wrong TOF, while the AIC picker of
the invention remained comparable to that selected by the manual
pick.
[0087] FIG. 4 illustrates an example of TOF pick overlays by the
AIC picker of the present invention on corresponding in vivo
ultrasonic breast waveforms. It should be noted that the solid dot
on each waveform segment indicates the TOF pick for that
waveform.
[0088] Tomograms obtained using improved wavelet-AIC picks
according to the present invention are compared with those selected
using amplitude threshold picks for in vitro and in vivo ultrasound
datasets.
[0089] FIG. 5A-5C illustrates a comparison between breast phantom
images.
[0090] The X and Y axes in these example images span 220 mm in
length. In FIG. 5A a representative X-ray CT scan of a breast
phantom is shown. In FIG. 5B is a tomogram obtained using the
improved AIC picks according to the invention. FIG. 5C illustrates
a tomogram obtained using amplitude threshold picks. It should be
readily recognized that the tomogram of FIG. 5B contains far fewer
artifacts than in FIG. 5C.
[0091] FIG. 6A and FIG. 6B illustrates a comparison of sound-speed
tomograms for in vivo ultrasound breast data in which TOF picks are
generated using TOF pick methods according to the present invention
in FIG. 6A and by utilizing amplitude threshold picks in FIG.
6B.
Section B
[0092] Section A provided a description and summarization of
aspects of the invention, while this section (Section B) describes
aspects of the invention in large part from the original
description. It should be appreciated that many of the equations
and figures used herein may duplicate those found in Section A.
Figure numbering is continued from Section A, but equation
numbering is restarted for section B to provide consistency with
the original text. Reference citation numbers are retained in this
section to provide additional information.
[0093] 7. Section B: Introduction.
[0094] Ultrasound sound-speed tomography has great potential to
detect and diagnose breast cancer. [1,2,3,4]. A clinical prototype
of ultrasound breast-imaging system with a ring array, termed the
Computed Ultrasound Risk Evaluation (CURE), has been developed at
the Karmanos Cancer Institute, Wayne State University in Detroit,
Mich. for ultrasound tomography [5 ].
[0095] FIG. 7A is a schematic illustration of a ring transducer
array performing ultrasonic imaging. In FIG. 7B a schematic of
Interaction of an ultrasound pulse with a target leads to a
scattered ultrasound field from transmitter (Tx) to receiver
(Rx).
[0096] In general, breast cancer has a higher sound-speed than the
surrounding breast tissue. A primary purpose of CURE is to
efficiently and reliably produce sound-speed images of the breast
for cancer detection and diagnosis. A potential sound-speed
reconstruction method for such a purpose is time-of-flight (TOF)
ultrasound transmission tomography. Accurate picking of TOFs of
ultrasound transmitted signals is an extremely important step to
ensure high-resolution and high-quality reconstruction of the
sound-speed distribution.
[0097] For each two-dimensional (2D) slice of ultrasound breast
data, each element of the CURE device acts as a transmitter as well
as a receiver, and all elements receive the scattered sound waves
when one element transmits.
[0098] CURE acquires 70-80 slices of ultrasound data for whole
breast imaging, resulting in a large volume of ultrasound data for
each patient. Therefore, it is not feasible to manually pick TOFs
of transmitted ultrasound data for sound-speed tomography because
manual picking is too time-consuming (.about.600,000 waveforms
needs to be analyzed for each patient). Accordingly, an automatic
TOF picker which can properly identify TOF provides an important
tool for ultrasound tomography, particularly for clinical
applications.
[0099] Different automatic TOF pickers have been developed, in
particular for use with geophysical applications to reconstruct the
internal structure of the Earth. The techniques used in these
devices fall into three general categories. The simplest method is
the amplitude threshold picker that applies an absolute value of
the threshold to the band-pass filtered signal. It is not
applicable for data having low signal-to-noise ratios (SNRs). A
variation is called "Short-Term-Average/Long-Term-Average
(STA/LTA)" method using the signal's envelope [6]. The second type
of auto-pickers utilizes a running window. Certain characteristics
are repeatedly calculated within successive sections of the time
series, producing a time dependent function. The TOF is usually
identified by an obvious change in the behavior of this function
([7,8]). The third type of auto-picker relies on using the
coherence characteristic between traces. One among these pickers
convolves a shifting reference waveform with the signal. The TOF of
the signal is determined when the measure of the quality of the
match is a maximum. This method assumes that the signal is
reasonably similar to the reference waveform. Several papers
describe this type of picker, including [9,10,11 ].
[0100] In 1951 Kullback and Leibler [12] proposed what is now known
as the Kullback-Leibler information criterion to measure the
information loss when approximating reality using recorded data. In
the 1970s, Akaike (cited in [13]) proposed a model selection
criterion, the Akaike Information Criterion (AIC), which relates
the maximum likelihood with the Kullback-Leibler information
criterion and minimizes the information loss during model
selection. Sleeman and Eck [14] applied the AIC and autoregressive
(AR) techniques to detect the TOFs of seismograms, and their TOF
picker is called AR-AIC picker. Autoregressive techniques are based
on the assumption that a waveform can be divided into locally
stationary segments as an AR process and the segments before and
after the TOF point are two different stationary processes. On the
basis of this assumption, the AR-AIC picker can be used to detect
the TOF of a seismogram by analyzing the variation in AR
coefficients. The AIC is usually used to determine the order of the
AR process when fitting a time series. When the order of the AR
process is fixed, the AIC is a measure of the model fit. In the
AR-AIC picker, the order of the AR coefficient is determined on a
trial and error basis (for details see [14]). To overcome this
difficulty and inefficiency, Zhang et al. [8] proposed a
wavelet-AIC picker in which the AIC values are calculated directly
from the seismogram using Maeda's formula [15]. In this method, a
running window and a wavelet transform are used to guide the AIC
picker by finding the appropriate time window that includes the TOF
point of a seismogram.
[0101] All the above techniques were historically developed to pick
elastic signals, particularly seismic waves. However, these
underlying concepts of these geologic mechanisms have not been
utilized for automatically picking TOFs for in vivo medical
ultrasound data. Kurz et al. [16] is one of the few who applied an
auto-picker to acoustic emission in concrete.
[0102] An improved AIC automatic TOF picker is taught according to
the present invention which is particularly well suited for in vivo
ultrasound breast data based on the wavelet-AIC TOF picker
described in [8]. The improved method makes use of an approach
termed multi-model inference (model averaging), based on the
calculated AIC values, to enhance the accuracy of TOF picks.
Aspects of the present invention also investigate applying a median
filter to remove TOF outliers. Demonstration of the inventive
automatic TOF picker shows that it can accurately pick TOFs in the
presence of random noise of up to 80% of the maximum absolute
synthetic signal amplitude. The improved automatic TOF picking
method is applied to clinical ultrasound breast data which
demonstrates that ultrasound sound-speed tomography with our
improved automatic TOF picks significantly enhances the
reconstruction accuracy while reducing image artifacts.
[0103] 8. Ultrasound Breast Data Acquired Using the CURE
Device.
[0104] The clinical ultrasound breast data used for this study was
collected with the CURE device, a clinical prototype ultrasound
scanner designed for clinical ultrasound breast tomography. CURE is
capable of recording all ultrasound wavefields including reflected,
transmitted, and diffracted ultrasonic signals from the breast
tissue. The engineering prototype of CURE is described in [17], and
the current clinical prototype is described in [5]. FIG. 7A-7B are
schematic representations of the transducer ring and Tx to Rx
pattern for a given pulse. FIG. 7B illustrates scattering of
ultrasound emitted from a transducer element and received by all
transducer elements along the ring. By way of example and not
limitation, there are a total of 256 elements in this 20-cm
diameter ring array. Each element is configured to emit and receive
ultrasound waves with a central frequency of 1.5 MHz. During the
scan, the ring array is immersed in a water tank, and encircles the
breast. The signals are recorded at a sampling rate of 6.25 MHz.
The whole breast is scanned slice by slice, and the scanned slice
data are recorded by a computer for data processing afterwards. A
motorized gantry is used to translate the ring along the vertical
direction, starting from the chest wall to the nipple.
[0105] 9. Improved Automatic AIC Time-Of-Flight Picker.
[0106] The wavelet-AIC TOF picker [8] is based on the AR-AIC picker
which assumes that a signal can be divided into locally stationary
segments and that the segments before and after the time-of-flight
point are two different stationary processes [14]. Data points
within the selected time window are divided into two segments at
each data point i(i=1, . . . , k, . . . N),where N is the total
number of data points in the selected time window. For a given data
point k the wavelet-AIC TOF picker uses Maeda's formula [15] to
calculate the AIC function directly from the waveform:
AIC(k)=k log (var(S(1, k)))+(N-k-1) log (var(S(k+1, N))) (1)
where S(1,k) (for data points 1 through k) and S(k+1, N) (for data
points k+1 through N) are the two segments in the selected time
window, and the variance function "var(.)" is calculated using
var ( S ( i , j ) ) = .sigma. j - i 2 = 1 j - i l = i j ( S ( l , l
) - S _ ) 2 , i .ltoreq. j , i = 1 , , N and j = 1 , , N ( 2 )
##EQU00004##
where S is the mean value of S(i,j).
[0107] The AIC value given by equation (1) measures the information
loss of using the current selected model to approximate reality. In
the wavelet-AIC auto-picker in [8], selecting the point with
minimum AIC value indicates the minimum information loss, wherein
it is called the best model, to be the TOF point.
[0108] The present invention provides improvements to automatic
wavelet-AIC TOF picking according to each of the following, to be
considered separately or in combination: (1) using a weighted
average model instead of the best model to determine TOFs; (2)
removing outliers of TOF picks using a median filter; and (3)
eliminating effects of signal distortion due to filtering and
wavelet de-noising during data preprocessing for improving SNR. A
schematic flowchart of the improved AIC TOF picker according to the
present invention has already been shown in FIG. 1 described in
Section A. The details of the improved AIC method of the current
invention are described in the following.
[0109] 9.1. Using a weighted average model to determine TOFs.
[0110] An AIC value by itself has no physical meaning and it
becomes interesting only when it is compared to a series of a
priori specified models [18]. The model with the minimum AIC value
is the best among all models being compared. The measure associated
with the AIC value that can be used to compare models is the
normalized Akaike weights (Eqs. 3 and 4). Akaike weights indicate
the relative importance of the candidate models. In most cases, the
best model (corresponding to the minimum AIC value) may have
competitors for the top rank. An elegant solution to make an
inference based on the entire set of models is to compute the
weighted average based on the model uncertainties (i.e. Akaike
weights). This is referred to model averaging or model
inference.
[0111] The running window used in for the wavelet-AIC TOF picker
[8] is not necessary for clinical ultrasound data. An appropriate
time window used for the TOF picks can be well-defined based on the
transmitter-receiver geometry and the sound-speed of water since
the sound-speed of breast tissue is close to that of water. To
incorporate all the information near the best model, a weighted
average model is utilized to pick the TOF in the following
sequence:
[0112] (1) Calculate AIC values (AIC.sub.i, i=1, . . . n) for a
series of data points near the point with the minimum AIC
(AIC.sub.min) using equation (1).
[0113] (2) Obtain the differences between AIC.sub.i(i=1, . . . n)
and AIC.sub.min:
.DELTA..sub.i=AIC.sub.i-AIC.sub.min (3)
[0114] (3) Compute the Akaike weights for each data point within
the time window:
w i = exp ( - .DELTA. i / 2 ) r = 1 n exp ( - .DELTA. r / 2 ) ( 4 )
##EQU00005##
[0115] (4) Obtain the TOF value using the weighted average:
t TOF = i = 1 n w i t i ( 5 ) ##EQU00006##
where t.sub.i (i=1, . . . n) are the corresponding travel times for
those points discussed in (1)-(3) and .omega..sub.i is obtained
using equation (4).
[0116] When there is a sharp global minimum, the AIC value that
indicates a high SNR, the difference between the TOF pick based on
the weighted average model and that based on the best model is
negligible. However, if the global minimum is not very sharp, which
indicates a low SNR, the weighted average model can pick the TOF
more accurately than picking based on the best model. This is one
of the only advantages of the automatic TOF picking method based on
the weighted model averaging scheme.
[0117] FIG. 8 illustrates an example of comparison among the TOF
pick based on the best model, the TOF pick based on the weighted
average model, and the manual pick of an in vivo ultrasound breast
signal acquired using the CURE device. The circle (".largecircle.")
represents the TOF picked with the best model, the cross ("X")
corresponds to the manual pick of the TOF, and the asterisk ("*")
is the TOF pick using the weighted average model. It can be seen
that the latter is closer to the manual TOF pick than the pick
based on the best model.
[0118] 9.2. Removing outliers of TOF picks.
[0119] To eliminate outliers of the TOF picks, a median filter is
applied to the time differences (TDs) between the TOFs picked
according to the present invention from ultrasound breast data and
calculated using the water sound-speed and the ring array geometry.
The median filter is a good tool for reducing `salt and pepper`
noise (outliers). To take advantage of this property and the
continuity property of a TOF surface (formed by TOFs of all
transmission data) ([19]), TDs are rearranged into a 2-D matrix
such that each row represents the TD values for a single
transmitter, and TD values for adjacent transmitters in the ring
array are put into adjacent rows (except the first and last
transmitter due to the circular geometry of the ring). This
rearrangement results in a 256 by 256 matrix(D). Another 256 by 256
matrix (M) containing all median values of TOFs is calculated with
a sliding window of the same size as the median filter. Adaptive
thresholds for the median filter are set up by calculating the
standard deviation (STD) and the mean value (ME) of TDs:
TolMin=ME-f*STD,
TolMax=ME+f*STD. (6)
where TolMin and TolMax are the minimum (could be a negative value)
and maximum tolerance for TDs, respectively, and f is a given scale
factor of the standard deviations with a value between 0 and 1. The
median filter based on the above thresholds is applied to the
matrix D: if D(i,j)-M(i,j)<TolMin or D(i,j)-M(i,j)>TolMax,
then the corresponding picked TOF is replaced with the median
value.
[0120] To further clean up the remaining picks, the TOF picks for
the reciprocal transmitter-receiver pair are compared against each
other. Reciprocal transmitter-receiver pair, here, means that two
transducers in the ring array transmit and receiver signals in the
opposite directions. Ideally, the TOF picks for the reciprocal pair
should be the same, which rarely happens in practice. In the
example data cleaning process, if the TOF's difference between the
reciprocal pair exceeds a predefined threshold, both picks are
discarded. The predefined threshold value can be customized by
users based on their individual requirement and data quality.
[0121] FIG. 9A-9B illustrates original matrix D and its
post-processed version for an in vivo ultrasound breast data
acquired using the CURE device. By way of this example both x and y
axes span 200 mm in length. Compared with FIG. 9A, FIG. 9B shows
that the inconsistent picks and outliers are effectively
eliminated.
[0122] FIGS. 9C-9D are ultrasound sound-speed transmission
tomography results for an in vivo ultrasound breast dataset using
the TOF picks shown in FIGS. 9A and 9B, respectively. These figures
demonstrates that the example TOF data cleaning procedure described
above can effectively remove TOF outliers and greatly improve the
quality of ultrasound TOF sound-speed tomography images.
[0123] 9.3. Signal distortion due to filtering and wavelet
de-noising.
[0124] Filtering and de-noising techniques are usually used to
preprocess a signal to improve its SNRs. The wavelet-AIC TOF picker
[8] applies the wavelet de-noising to a raw seismogram before it
picks the TOF. In fact, both the filtering and wavelet de-noising
may distort a signal while attempting to increase the SNR [20
].
[0125] FIG. 10A-10C illustrates a comparison of a raw ultrasound
data segment (FIG. 10A) acquired with the CURE device with its
filtered version (FIG. 10B) and wavelet de-noised version (FIG.
10C). The signal in FIG. 10B was filtered using a second-order
zero-phase Butterworth band-pass filter with the stop band corner
frequencies at 0.3 MHz and 2.3 MHz, and the pass band corner
frequencies at 0.9 MHz and 1.7 MHz. For the wavelet de-noised
signal in FIG. 10C, thresholding was applied to the wavelet
coefficients using the Birge-Massart penalization method [21]. The
solid vertical line in each of the figures represents the picked
TOFs from the raw ultrasound data, while the dashed line indicates
the picked TOFs from the filtered and de-noised segments.
[0126] Coincidently, TOFs picked from the band-pass filtered signal
and the wavelet de-noised signal are the same as seen in Table 1.
From FIG. 10A-10C signal distortions due to the wavelet de-noising
and zero-phase band-pass filtering can be seen. Because of the
short transient time of ultrasound signals, a small distortion of
the ultrasound waveform may result in large unwanted artifacts or
erroneous information during TOF picking. For these reasons and in
order to preserve the true shapes of ultrasound signal onsets as
much as possible, no signal preprocessing is performed toward
improving the SNRs, because the improved automatic TOF picker of
the invention can handle ultrasound data with low SNRs as
demonstrated in the following.
[0127] 9.4. Capability to handle noisy data.
[0128] To estimate the maximum amount of random noise that our
improved wavelet-AIC TOF picker can tolerate, progressively greater
amounts of random noise were added to a synthetic ultrasound
waveform with a similar spectrum to that of in vivo ultrasound
breast data acquired by CURE. The average absolute amplitudes of
the random noise were, respectively, 0%, 20%, 40%, 60% and 80% of
the maximum absolute amplitude of the synthetic ultrasound signal.
The improved TOF picker according to the present invention can
consistently detect the correct TOF in the presence of 80% white
noise, which corresponds to a 4.5 dB SNR.
[0129] FIG. 11A-11B and 12A-12 B illustrate signals and
corresponding AIC values for both a high and low SNR ultrasound
signal. In FIG. 11A and 12A are seen example signal representations
with 0% and 80% random noise added, respectively. FIG. 11B and 12B
indicate the value of AIC, from which a TOF can still be selected
even at 80% noise added (SNR: 4.5 dB). The tests conducted herein,
as well as work on preceding forms of AIC, such as the work of Kurz
in [16] show that AIC-based TOF pickers can tolerate a relatively
high noise level if an appropriate time window is used.
[0130] 10. Assessment of Inventive AIC Time-Of-Flight Picker.
[0131] To assess the performance of the improved AIC picker, TOF
picks according to the present invention were compared with those
produced through an amplitude threshold picker, and via manual
picking; in this case of 1160 waveforms of in vivo ultrasound
breast data acquired using the CURE device. Manual picking was
conducted by recognizing the first rise time of the signal. To
exploit the continuity of picked TOFs for adjacent waveforms [19],
five consecutive waveforms were plotted on the computer monitor at
the same time to further improve the accuracy of manual
picking.
[0132] FIG. 13 is a snapshot image of the manual picking process,
in which solid triangles indicate the manual TOF picks which were
selected in response to clicking the computer mouse. To better
illustrate ultrasound waveforms, the time windows were selected
from 45 .mu.s to 100 .mu.s.
[0133] Since the amplitude threshold picker is much more sensitive
noise, band-pass filtering was applied before picking the TOFs
using an amplitude threshold. A second order zero-phase Butterworth
band-pass filter with stop band corner frequencies at 0.3 MHz and
2.3 MHz, and pass band corner frequencies at 0.9 MHz and 1.7 MHz
was used to filter the ultrasound breast data. To make a fair
comparison, the same outlier removal procedures (median filtering
and reciprocal pair check) were also applied to the amplitude
threshold TOF picks.
[0134] FIG. 14A-14B illustrate time difference data for 1160 in
vivo ultrasound breast waveforms acquired using the CURE device (a
ring transducer array). FIG. 14A illustrates the absolute values of
TOF differences between manual picks and amplitude threshold picks,
while FIG. 14B illustrates those TOF difference between manual
picks and improved AIC picks according to the present invention.
The statistics are given in Table 2 showing that for these 1160 in
vivo ultrasound breast waveforms, over 85% of the TOFs picked by
the inventive AIC picker are within three sample points (0.48
.mu.s) from the manual picks. The mean value and standard deviation
between the inventive TOF picker and manual picking are 0.4 .mu.s
and 0.29 .mu.s, respectively. Among picks by the amplitude
threshold picker, only 48% of them are within the three sample
points from the manual picks, and the mean value and standard
deviation from the manual picks are respectively 1.02 .mu.s and 0.9
.mu.s, which are much higher than those of the inventive TOF
picking method.
[0135] Studies performed in association with this application
utilizing other clinical ultrasound breast data indicate that the
performances of our improved wavelet-AIC TOF picker and amplitude
threshold-based TOF picker are comparable for data with high SNRs,
but for data with high noise level (or low SNRs), the accuracy of
the amplitude threshold picker drops abruptly. Moreover, for noisy
ultrasound breast data, the failure rate of the amplitude threshold
TOF picker is much higher than that of the improved wavelet-AIC TOF
picker of the invention.
[0136] FIG. 9A-9B illustrates two waveforms of different SNRs along
with TOF picks by the above three methods. The type of TOF pick is
shown in the figure based on best model (".largecircle."), manual
pick ("X"), and weighted average ("*") according to the present
invention. It can be clearly seen that for the waveform in FIG. 9A,
which has a high SNR, the three TOF picks are consistent. The AIC
pick according to the invention is closer to the manual pick than
is the best model (amplitude) pick, while in response to a noisy
waveform, such as in FIG. 9B, the amplitude threshold picker
(".largecircle.") picked the wrong TOF, while the improved
wavelet-AIC TOF pick according to the present invention remains
comparable to the manual pick. The TOF picks in FIG. 9A-9B are
shown in Table 3.
[0137] FIG. 16 illustrates an example of overlays of the TOF picks
by the inventive AIC TOF picker on the corresponding in vivo
ultrasonic breast waveforms in which the solid dot on each waveform
segment indicates the inventive AIC based TOF pick.
[0138] 11. Tomography Results of in vitro and in vivo Ultrasound
Data.
[0139] TOF sound-speed tomography uses the TOF picks of ultrasound
breast data for reconstruction. The tomogram quality depends
directly on the quality and accuracy of TOF picks.
[0140] FIG. 17A-C compares tomograms obtained using X-ray CT scans
with ultrasonic imaging obtained for in vitro and in vivo
ultrasound datasets using the inventive AIC TOF picks in comparison
with amplitude threshold TOF picks. For this example all x and y
axes span 200 mm in length. FIG. 17A is a cross-section image from
an X-ray CT scan of a breast which includes a phantom (tumor). The
ultrasound sound-speed tomograms obtained using inventive AIC TOF
picker is shown in FIG. 17B, while the tomogram for using amplitude
threshold TOF picks is shown in FIG. 17C. It can be seen that the
tomogram in FIG. 17B contains significantly fewer artifacts than
the tomogram in FIG. 17C and much more closely resembles that of
the X-ray CT scan. Moreover, the four inclusions and surrounding
subcutaneous fat are better reconstructed in FIG. 17B than those in
FIG. 17C. In contrast to in vivo ultrasound breast data, the
phantom breast data has relatively low structure noise due to
relatively simple internal structures.
[0141] FIG. 18A-18B illustrate another comparison for in vivo
ultrasound breast data acquired with the CURE device. For this
example all x and y axes span 100 mm in length. The in vivo breast
data used to obtain these images has a relative low SNR (.about.18
dB). The sound-speed tomogram produced using the inventive TOF
picks, of FIG. 18A, appears to have fewer straight-line artifacts
compared to the one obtained using amplitude threshold TOF picks as
shown in FIG. 18B. For in vivo ultrasound breast data with relative
high SNR (.about.25 dB), the difference between the corresponding
sound-speed tomograms is minor, although the reconstruction with
the inventive TOF picking still appears to be superior to the one
produced using amplitude threshold TOF picks in terms of the mass
detection and reconstruction noise. These comparisons again
demonstrate that improved AIC TOF picking according to the present
invention is much less sensitive to varying SNRs of ultrasound data
than the amplitude threshold TOF picker.
[0142] 12. Section B: Conclusions.
[0143] This section has described an inventive automatic TOF
picking method based on the Akaike Information Criterion (AIC)
which has been successfully applied to in vivo ultrasound breast
data, which for example has been collected using a ring transducer
array. To improve the accuracy of the TOF picking, the improved
picking method according to the invention incorporates all the
information near the TOF point using a model inference method to
determine the TOF of ultrasound signals. Further aspects of the
invention utilize a median filter to remove TOF outliers. The
resultant TOF picker of the invention can pick correct TOFs in
noisy ultrasound data (with average absolute amplitudes of noise up
to 80% of the maximum absolute amplitude of the signal) while the
amplitude threshold based TOF picking method generally fails. For
ultrasound breast data, the present inventive method is thus able
to determine TOFs of a similar accuracy (quality) to those picked
manually by an expert. One of the important advantages of the
automatic TOF picker according to the invention is that it is
operator independent, and is much less time-consuming than the
manual picking. Accordingly, the present method makes it possible
to incorporate automatic TOF picking within a clinical ultrasound
tomography device without sacrificing outcome quality. It has been
demonstrated that ultrasound sound-speed tomography using TOFs
picked using the inventive automatic TOF picker method
significantly improves the reconstruction accuracy and reduces
image artifacts.
Section C
[0144] It should be appreciated that although the methods described
were directed at ultrasonic breast tomography, these techniques can
be implemented within any number of ultrasonic tissue imaging
apparatus. The method is particularly well-suited for
implementation on a system which receives ultrasonic waveforms and
utilizes a computer for processing those signals. It should be
appreciated, however, that the aspects of the invention can be
implemented on any desired combination of software and hardware as
will be recognized by one of ordinary skill in the art.
[0145] FIG. 20 illustrates an embodiment 50 of an ultrasonic
imaging apparatus according to the present invention. A sensor head
52 is shown exemplifies as a ring configured for breast tomography,
although it can be configured in any desired configuration for
various forms of tissue testing. The sensor head 52 is configured
with transmitters and receivers controlled by block 54. All
necessary data from the sensor head is conditioned as necessary in
signal conditioning block 56, from which data 58 on a plurality of
ultrasonic signals is communicated to a computing device 60
containing at least one processing element 62 and memory 64.
Programming executable on computer 62 is configured for retention
in memory 64, and for executing the described method steps
according to the present invention. The TOF picks can be utilized
internal to the computer or be output 66 from the computer for use
by image processing equipment 68 and image display and/or storage
elements 70. It will thus be appreciated that numerous medical
ultrasonic devices can be configured according to the teachings of
the present invention to improve resolution and quality of the
ultrasonic information.
[0146] Although the description above contains many details, these
should not be construed as limiting the scope of the invention but
as merely providing illustrations of some of the presently
preferred embodiments of this invention. Therefore, it will be
appreciated that the scope of the present invention fully
encompasses other embodiments which may become obvious to those
skilled in the art, and that the scope of the present invention is
accordingly to be limited by nothing other than the appended
claims, in which reference to an element in the singular is not
intended to mean "one and only one" unless explicitly so stated,
but rather "one or more." All structural and functional equivalents
to the elements of the above-described preferred embodiment that
are known to those of ordinary skill in the art are expressly
incorporated herein by reference and are intended to be encompassed
by the present claims. Moreover, it is not necessary for a device
or method to address each and every problem sought to be solved by
the present invention, for it to be encompassed by the present
claims. Furthermore, no element, component, or method step in the
present disclosure is intended to be dedicated to the public
regardless of whether the element, component, or method step is
explicitly recited in the claims. No claim element herein is to be
construed under the provisions of 35 U.S.C. 112, sixth paragraph,
unless the element is expressly recited using the phrase "means
for."
[0147] 13. References for Section B.
[0148] [1] J. F. Greenleaf, A. Johnson, R. C. Bahn, and B.
Rajagopalan, Quantitative cross-sectional imaging of ultrasound
parameters, in Proc. IEEE Ultrason. Symp. (1977) 989-995.
[0149] [2] S. J. Norton and M. Linzer, Ultrasonic reflectivity
tomography: reconstruction with circular transducer arrays,
Ultrason. Imag 2 (1979) 154-184.
[0150] [3] P. L. Carson, C. R. Meyer, A. L. Scherzinger, and T. V.
Oughton, Breast imaging in coronal planes with simultaneous pulse
echo and transmission ultrasound, Science 214 (1981) 1141-1143.
[0151] [4] M. P. Andre, H. S. Janee, H. S., P. J. Martin, G. P.
Otto, B. A. Spivey, and D. A. Palmer, High-speed data acquisition
in a diffraction tomography system employing large-scale toroidal
arrays, Int. J. Imaging Syst. and Technol. 8 (1997) 137-147.
[0152] [5] N. Duric, and P. Littrup, L. Poulo, A. Babkin, R.
Pevzner, E. Holsapple, and O. Rama, Detection of Breast Cancer With
Ultrasound Tomography: First Results with the Computerized
Ultrasound Risk Evaluation (CURE) Prototype, Med. Phys. 34 (2007)
773-785.
[0153] [6] M. Baer and U. Kradolfer, An automatic phase picker for
local and teleseismic events, Bull. Seism. Soc. Am. 77 (1987)
1437-1445.
[0154] [7] F. Boschetti, D. Dentith, and R. D. List, A
fractal-based algorithm for detecting first-arrivals on seismic
traces, Geophysics 61 (1996) 1095-1102.
[0155] [8] H. Zhang, C. Thurber, and C. Rowe, Automatic P-wave
Arrival Detection and Picking with Multiscale Wavelet Analysis for
Single-Component Recordings, Bull. Seism. Soc. Am. 93 (2003)
1904-1912.
[0156] [9] R. Ramananantoandro and N. Bernitsas, A Computer
Algorithm for Automatic Picking of Refraction First-arrival-time,
Geoexploration 24 (1987) 147-151.
[0157] [10] W. Su and A. M. Dziewonski, A. M., On the scale of
mantle heterogeneity, Phys. Earth Planet. Interiors 74 (1992)
29-54.
[0158] [11] J. B. Molyneux and D. R. Schmitt, First-break timing:
Arrival onset times by direct correlation, Geophys. 64 (1999)
1492-1501.
[0159] [12] S. Kullback and R. A. Leibler, On information and
sufficiency, Annals of Mathematical Statistics 22 (1951) 79-86.
[0160] [13] D. R. Anderson, K. P. Burnham, and G. C. White,
Kullbak-Leibler information in resolving natural resource conflicts
when definitive data exist, Wildlife Society Bulletin 29 (2001)
1260-1270.
[0161] [14] R. Sleeman and T. van Eck, Robust automatic P-phase
picking: an on-line implementation in the analysis of broadband
seismogram recordings, Phys. Earth Planet. Interiors 113 (1999) pp.
265-275.
[0162] [15] N. Maeda, A method for reading and checking phase times
in autoprocessing system of seismic wave data, Zisin. Journal of
Seismological Society of Japan 38 (1985) 365-379.
[0163] [16] J. H. Kurz,, C. U. Grosse, and H. W. Reinhardt,
Strategies for reliable automatic onset time picking of acoustic
emissions and of ultrasound signals in concrete, Ultrasonics 43
(2005) 538-546.
[0164] [17] N. Duric, and P. Littrup, A. Babkin, D. Chambers, S.
Azevedo, R. Pevzner, M. Tokarev, E. Holsapple, O. Rama, and R.
Duncan, Development of ultrasound tomography for breast imaging:
Technical assessment, Med. Phys. 32 (2005) 1375-1386.
[0165] [18] M. J. Mazerolle, Making sense out of Akaike's
Information Criterion (AIC): its use and interpretation in model
selection and inference from ecological data [Online]. Available:
http://www.theses.ulaval.ca/2004/21842/apa.html.
[0166] [19] C. Li and R. L. Nowack, R. L, Seismic Tomography Using
Travel-Time Surfaces for Experiments in the Laboratory, J. Geophys.
Eng. 2 (2005) 231-237.
[0167] [20] R. Di Stefano, F. Aldersons, E. Kissling, and C.
Chiarabba, Automatic seismic phase picking and consistent
observation error assessment: application to the Italian
seismicity, Geophys. J. Int. 165 (2006) 121-134.
[0168] [21] L. Birge and P. Massart, From model selection to
adaptive estimation, in Festchrifft for L. Le Cam, D. Pollard
Editor, Springer, New York, 55-88, 1997.
TABLE-US-00001 TABLE 1 TOF Picks for Different Conditions of FIG.
10A-10C Condition TOF (.mu.S) Raw data 100.80 Zero-Phase band-pass
filtered data 100.16 Wavelet de-noised data 100.16
TABLE-US-00002 TABLE 2 Comparison of AIC and Amplitude Threshold
Picks with Manual picks Std. Dev. of Mean Difference(.mu.S)
Difference (.mu.S) Improved AIC picks 0.4 0.29 Amplitude threshold
picks 1.02 0.9
TABLE-US-00003 TABLE 3 TOF picks by the Three Methods of FIG.
15A-15B FIG. 15A (.mu.S) FIG. 15B (.mu.S) Improved AIC picks 137.60
69.59 Manual picks 137.63 69.60 Amplitude threshold picks 138.16
72.07
* * * * *
References