Content uploaded by Lori Lee Barr
Author content
All content in this area was uploaded by Lori Lee Barr on Mar 14, 2019
Content may be subject to copyright.
Understanding the
Process of Quantitative
Ultrasonic Tissue
Characterization1
Janice W. Allison, MD #{149}Lori L. Barr, MD #{149}Ricbardj Massoth, PhD
Greg P. Berg, MS #{149}Brian H. Krasner, PhD #{149}Brian S. Garra, MD
Because the human vision system cannot distinguish the broad range of gray
values that a computer visual system can, computerized image analysis may be
used to obtain quantitative information from ultrasonographic (US) real-time
B-mode scans. Most quantitative US involves programming an off-line comput-
er to accept, analyze, and display US image data in a way that enhances the de-
tection of changes in small-scale structures and blood flow that occur with dis-
ease. Common image textural features used in quantitative US tissue charac-
tenization consist of first-order gray-bevel statistics (eg, occurrence frequency
of gray levels independent of location or spatial relationship) and second-order
gray-level statistics dependent on location and spatial relationship, including
statistical analysis of gradient distribution, co-occurrence matrix, covariance
matrix, run-length histogram, and fractal features. A customized tissue signa-
tune software has been developed to analyze image data obtained from clinical
Us scanners. Means comparison testing and multivaniate analysis techniques
are used to compare the numbers generated for a particular region of interest.
By integrating these techniques into the radiologist’s interpretation of the
sonogram, the quantitative information gained may lead to earlier detection of
lesions difficult to see with the human eye.
UINTRODUCTION
Qualitative interpretation of a sonogram requires visual recognition of patterns from
processed ultrasonographic (US) images by a skilled sonobogist. Quantitative evalua-
tion can be achieved if the analog or digital data from which the real-time B-mode scan
is derived are channeled into a computer system designed to analyze patterns in the
data. Much of the tissue information available in the radio-frequency (RF) data is dis-
carded by the current generation of clinical US instruments. Unless the manufacturer
provides access to the RF data, it is necessary to use envelope-detected image data
from clinical US instruments for off-line quantitative tissue characterization.
Abbreviations: RF radio frequency. ROl =region of intercst
Index terms: tJItr.oound (115) #{149}tiltrjsound (US), technology #{149}tIltr4sound (tTS). tissue characterization
RadioGraphics 1994; 14: J099- I 108
IIroni the 1)cpartrnents of Radiology. Arkansas Childrens hospital and tniversity of Arkansas for Medical Sciences, Little
Rock (J.W.A.); Childrens 1-lospital Medical Center. 3333 Burnet Ave. Cincinnati, OH 45229-3039 (L.L.B.): 1)ivision of Medical
Physics. I iniversity of Cincinnati. College of Medicine. Ohio (R.J.M.): and Georgetown I niversity Medical Center, Washing-
ton, I)C; (BIlK.. 13.5G.) and the t)epartment of Science. Maine Township High School East, Park Ridge. Ill (G.P.B.). Present-
ed as a scientific exhibit at the 1993 RSNA scientifIc assembly. Received April 5. 1994: revision requested May 24 and i-c.
ceivedjune 23; acceptedJune 29. Address reprint requests to L.L.B.
RSNA, 1994
1099
1100 #{149}Imaging & Therapeutic Technology Volume 14 Number 5
Figure 1. First- versus second-order visual tasks. (a) Coronal sonogram of the brain in an infant horn prenia-
turely. Recognizing the bright intrapanenchymal hemorrhage (arrows) is a first-order visual task easily per-
formed by sonologists. (b) Coronal sonogram of the brain in an infant with low Apgar scores at birth who re-
quired resuscitation. Recognizing subtle changes in the echotextune, such as coarsening caused by hypoxic-
anoxic insult (arrows), is a second-order visual task. Intra- and interobserver variability is greater when the
sonologist is trying to answer questions pertinent to the child’s treatment, such as what percentage of the brain
is affected or how severe is the injury?
This article presents an overview of the con-
cepts and history of ultrasonic tissue character-
ization, describes an example in which textural
data were used to aid the diagnosis of pediatric
brain abnormalities, and discusses textural fea-
tune analysis.
.US AS A DIAGNOSTIC TOOL
US is currently used in many areas of medicine
for probing tissues noninvasively, in real time,
at low cost, and at low risk to both patient and
sonographer. These attributes have established
sonography as an essential element of an effec-
tive diagnostic imaging program.
Conventional B-mode US images are formed
by transmitting an acoustic pulse into the body,
followed by the receipt, processing, and dis-
play of the amplitude of returning echo signals.
These signals are displayed as a map of ampli-
tude as a function of echo delay for different
acoustic lines of sight. The echo signals used to
form B-mode images reflect three chanacteris-
tics of the tissues: large-scale anatomic bound-
aries, small-scale internal structure, and motion
(eg, blood flow).
Unfortunately, not all of this information is
detectable with the human eye on either a hard
copy or real-time B-mode image. Sonograms ap-
pear speckled because they display both quasi-
regular structures, such as organ microarchi-
tectune, and interfering waves scattered from
very fine, diffuse, randomly positioned struc-
tunes, such as tiny blood vessels. Acoustic
speckle obscures diffuse parenchymal process-
es from qualitative detection. Pathologic enti-
ties such as hepatic cirrhosis and hypoxic or
anoxic injury in the brain result in changes in
the echotexture that may or may not be distin-
guishable from normal appearances in a qualita-
tive interpretation.
Figure 2. Embedded lattice technique of visual testing. (a) A grid is positioned in a quadrant of this image hut
is obscured by noise and is almost impossible to detect with the unaided eye. (b) The grid is positioned in the
lower left quadrant of the test diagram. (Fig 2a and 2b reprinted, with permission, from reference 4.)
September 1994 Allison et al #{149}RadioGraphics U1101
To observe and quantify small-scale struc-
tunes and blood flow directly at US, signal pro-
cessing beyond that used in conventional imag-
ing is necessary. This has prompted investiga-
tons to search for other ways to process the
echo signals and to obtain quantitative ultrason-
ic data from tissues (1,2). One such method is
textural feature analysis, which quantitates the
image texture caused by the acoustic speckle
in the tJS images. This method may be em-
pboyed in the clinical setting as a compliment
to qualitative interpretation.
. Qualitative US
Qualitative IJS involves the performance of
first- and second-order visual tasks by trained
human observers (Fig 1). Although the human
eye can be trained to become quite efficient at
first-rder visual tasks, there is a wide range of
human variability present in performance of
second-order tasks. Several investigations have
compared human visual system performance at
first- and second-order tasks with that of a com-
puten program to measure various acoustic pa-
rameters quantitatively (2,3). Figure 2 illus-
trates the embedded lattice technique of visual
testing (4). In these investigations, the comput-
en greatly outperformed humans at both first-
and second-order visual tasks. Although the hu-
man eye can appreciate pixel intensities on a
cathode nay tube display corresponding to only
1 5-30 shades of gray (5), the computer can dis-
tinguish as many shades of gray as are stoned in
its memory.
Two factors limit the qualitative interpneta-
tion of conventional B-mode images: (a) oh-
scuration of small-scale quasinegulan structures
by acoustic speckle and (b) human inefficiency
in performing second-order visual tasks. Practi-
cally speaking, qualitative US is easy to perform
in the clinical setting, whereas quantitative
methods have had only limited clinical use to
date.
1102 #{149}Imaging & Therapeutic Technology Volume 14 Number 5
Comparison o fRaw versus Processed Data for Quantita five Analysis of US Images
Raw Data Processed t)ata
Parameters Attenuation coefficient, speed of sound,
backscatten coefficient, first-order
scatterer size and strength statistics
Attenuation coefficient, speed of sound,
backscatten coefficient, first-order and
second-order gray-level statistics
Requirements Expensive high-speed digitizer, custom
electrical interface, machine calibration Access to digital image data, machine cali-
bration on repeatable scanning parameters
Advantage Parameters derived correlate with tissue Possible on most commercial units
structure and composition
Disadvantage Massive data stream Assumes that pixel intensities are related to
tissue structure and composition
.Quantitative US
Most quantitative US involves programming an
off-line computer to accept, analyze, and dis-
play ultrasonic image data in a way that en-
hances the detection of changes in small-scale
quasiregular structures and blood flow that fre-
quently occur with disease (2). The develop-
ment of pulsed Doppler velocity on amplitude
displays and color Doppler maps illustrates the
application of additional data manipulation to
quantify the velocity of blood flow. This type
of quantitative analysis, once performed off-
line, has become an integral part of every US
scanner used for color Doppler and color flow
imaging. It is likewise possible to acquire and
manipulate B-mode data to quantitate changes
in small-scale structures within an organ (6,7).
This latter technology is not currently available
on standard commercial systems.
In general, quantitative ultrasonic tissue
characterization aims to measure or estimate
acoustic properties, such as backscatter, alien-
uation, on sound speed (2), that determine the
imaging appearance of tissues. These measur-
able acoustic properties depend on cell size,
the material between cells, and the arrange-
ment of the cells, which affect the viscoelastic
properties of the tissue: density, compnessibii-
ty, and the acoustic absorption coefficient
(8,9). On the basis of the assumption that dis-
ease processes alter the normal tissue structure
and viscoelasticity, quantifying the measurable
acoustic properties provides important diag-
nostic information beyond what is accessible
by qualitatively observing the image.
. Tissue Characterization
The first step in the process of tissue character-
ization is the capture of the ultrasonic RF data.
US waves may be captured after they either
pass through on are reflected back by tissue. RF
data collected after the sound waves pass
through tissue are called through-transmission
data. The majority of studies using through-
transmission data have been performed in laho-
ratory settings, in which tissues are isolated
and embedded in agan and then suspended in a
water bath. Transmitting and receiving trans-
ducers are positioned on opposite sides of the
sample. Hand-held through-transmission instru-
ments are shaped bike calipers, which limits ap-
pbications in vivo to small parts or intraopera-
tive imaging.
The majority of in vivo measurements in-
volve collecting RF data reflected hack from tis-
sue (backscattered data). The acoustic data
needed for off-line quantitative analysis may he
captured at several points in the conventional
image formation process: (a) as they return to
the transducer (raw RF data), (b) after prepro-
cessing (envelope-detected data). or (c) after
postprocessing (digital image data, hand-copy
image, or video output data).
Noise increases during amplification of the
returning signal. Electronic filters such as band-
pass filters remove much of this noise, thus im-
proving the tissue characterization process.
Some electronic filters may increase the uncer-
tainty of derived image parameters, since noise
may also increase. Timing jitter in the video
output or video capture (frame-grabbing) pro-
cess may cause a blurring (time averaging) or
uncertainty (“jumps” or “glitches”) of data ac-
quined in this manner; therefore, video capture
September 1994 Allison et al #{149}RadioGraphics #{149}1103
is not recommended. The Table enumerates
the calculable parameters and limitations of
quantitative analysis with raw versus pro-
cessed data.
UHISTORIC TISSUE CHARACTERI-
ZATION
Ultrasonic tissue characterization has been
used to analyze normal and diseased tissues for
more than 20 years. Although tissue character-
ization has been attempted in a number of ana-
tomic structures, including the liver (10-16),
spleen (16,17), kidneys (6,7,18), breast
(19,20), muscle (21-23), and eye (24-26), the
areas that have received the most attention
have been the liver and the eye. The liven is a
popular structure for evaluation because it is
affected by a number of diffuse diseases and is
a large organ. This allows for collection of data
from barge regions of interest (ROIs), thus re-
ducing variability in the quantitative results.
A variety of tissue characterization tech-
niques have been attempted. These fall into
two major categories: those that analyze the
statistics of the image data and those that try to
compute tissue parameters from the RF signal.
Initially, much attention was focused oii the
computation of the attenuation coefficient in
tissues. Several methods were developed to
compute this quantity, but the results were
never very clinically useful because most liver
diseases produce only minimal effects on alien-
uation. Fatty infiltration of the liver produces
the largest effect on attenuation, but it is not
an entity of great clinical importance. Attenua-
tion computation in muscle tissue, however,
has been useful in the differentiation between
muscles involved with Duchenne (pseudo-
hypertrophic) muscular dystrophy and healthy
muscles (23). Numerous attempts have been
made to compute the speed of sound in vari-
0U5 tissues, but this measurement alone has
not been very useful in categorizing various
disease states, either in the liven on the spleen.
For liver tissue characterization, the most
useful features have been those that depend on
analysis of the matrix of image data and those
derived from the power spectrum of the auto-
correlation of image data (1 5). In the eye and
the kidney, features derived from analysis of
the RF dependence of the backscatter coeffi-
cient have been found to be useful in both the
detection of disease and the differentiation be-
tween various disease states. In the kidney, the
frequency dependence of the backscatten coef-
ficient can be used to estimate gbomeruban size
(6,7). In the eye, metastatic melanoma in the
retina can be distinguished from other tumor
masses (24,25). In the breast, both features de-
rived from frequency-dependent backscatter
coefficients and those from image statistics
have proved useful in the differentiation of
one type of mass from another.
The problem with all of these techniques
has been that they are difficult to implement
on standard commercial systems and often pro-
duce numeric results rather than an image.
Evaluation of tissue elasticity or hardness is a
promising new approach that is just now be-
ing applied to breast lesions, renal transplants,
and the prostate gland. This method does pro-
duce an image (27). The advent of commercial
workstations for viewing and analyzing images
and of tissue characterization software for
these stations may bead to wider use of the
techniques.
U QUANTITATWE ANALYSIS IN THE
CLINICAL SE1TLNG
One example of integration of quantitative
techniques into the clinical and laboratory set-
ting is described herein. We currently perform
off-line tissue characterization on processed
digital image data. In some experiments, the
data were downloaded from the US scanner
(model UM-9; Advanced Technology Laborato-
ries, Bothell, Wash) into a computer (Macin-
tosh IIfx; Apple Computer, Cupertino, Calif)
by using an RS-422 serial port and proprietary
software provided by the manufacturer. Using
the Aegis System (Acuson Computed Sonogra-
phy, Mountain View, Cabif), we stone digital
image data files from an Acuson ART US scan-
ncr. Through machine calibration or the use of
a calibration phantom, the variability produced
by different machine parameters, such as trans-
ducer frequency, image depth, and focus, may
be measured. After scan conversion, the im-
ages are transferred to a workstation (SPARC-
station 2; Sun Microsystems, Mountain View,
Calif), where textural analysis is performed
with a custom software package developed for
the Sun operating system.
- ,#{149}. ,..-, I.
The customized software allows image visu-
alization, ROI selection, and textural feature
analysis (Fig 3). For images of the brain, we
presently employ a 30 x 30- or 40 x 40-pixel
ROL. The 30 x 30-pixel ROI corresponds to a tis-
sue volume of 0.9 mL, with a 10-MHz transduc-
en set used to a maximum depth of 3 cm (a typi-
cab laboratory setting). The textural feature data
may be transferred in either numeric or graphic
file formats to a Macintosh computer for gener-
ation of comparative data bases and statistical
analysis. Means comparison testing and multi-
variate analysis techniques are used to compare
the numbers generated for groups of ROIs.
Cl)
C)
0.
0
L.
C)
E
z
0246810 12
Gray Level
1104 UImaging & Therapeutic Technology Volume 14 Number 5
-
., ,i,l
ILI ii:iii
___________________________________1ff:!.. J, ‘:; .IT....: :. : _______
I
Figure 3. Sun SPARCstation 2 workstation screen with the tissue signature software operating. Photograph
shows a selected 8-bit image, with the selected size and position of the ROI. Clicking on “Features” (arrow)
starts the calculation of textural parameters. The results are displayed in the boxes shown and are recorded in a
text file for easy export into a data base.
Figure 4. Graph of first-order gray-level statistics
shows the occurrence frequency of ROI pixel inten-
sities.
September 1994 Allison et al U RadioGraphics #{149}1105
Figures 5, 6. (5) Gradient analysis. I)iagram depicts a 5 x S-pixel ROI. Gradient
analysis (seconthrder gray-level statistics) begins by defining a neighborhood for
each pixel. (ray-level differences are determined by comparing the intensity of pixel
M with those of its neighbors. When this is performed for each pixel in the ROI, the
gradients may he determined. (6) Run-length histogram analysis. Diagram depicts a 5
xS-pixel ROt with a horizontal run of 3 pixels length left of pixel M. In run-length
histogram analysis, the sets of pixels with the same gray level are counted.
.TEXTURAL FEATURE ANALYSIS
TECHNIQUES
Texture measurements that were useful in the
past are those that identify tissue signatures.
These parameters were first described in the
broaden field of digital signal processing and
have since been applied to medical US (28-3 1).
The first- and second-order gray-level statistics
are based on the pixel intensities found in the
selected ROl. Finst-nder gray-level statistics in-
elude the occurrence frequency of the gray 1ev-
els independent of location on spatial relation-
ship (Fig 4). The gray-level histogram allows cal-
culation of mean gray level, gray-level variance,
and skewness (deviation from symmetry). The
first-order gray-level statistics correspond to the
overall echogenicity of an anatomic area as per-
ceived by the sonologist. The second-order
gray-level statistics are dependent on location
OI spatial relationship and include statistical
analysis of gradient distribution, co-occurrence
matrix, covariance matrix, run-length histo-
gram, and fractal features.
In gradient analysis, the difference in intensi-
ties between cacti pixel and all neighboring pix-
els is evaluated (32). The gradient is calculated
by defining a rectangular neighborhood for
each pixel (Fig 5) and then calculating the abso-
lute value and direction of the local gray-level
difference gradient. Mean gradient absolute val-
ue, gradient value variance, and relative fne-
quency of the most dominant edge with respect
to the number of gradient elements in the ROI
may then be calculated. Gradient analysis come-
sponds to edge detection by the sonologist.
Co-occurrence matrix analysis involves gen-
eration of a two-dimensional histogram charac-
terizing the occurrence of gray-level combina-
tions in spatially rebated pixel pairs (32). Analy-
sis of the two-dimensional histogram involves
the search for a pattern of intensity variation
between a pixel and its neighbors in each car-
dinal direction. The neighboring pixels may be
closely (2 pixels apart) on widely (4 pixels
apart) spaced. After assignment of the various
directions, such as east-west, north-south, or
northwest-southeast, the following characteris-
tic features are extracted: (a) contrast, a mea-
sure of how many large gray-bevel differences
are present in the ROI; (b) angular second mo-
ment, a measure of the degree of clustering of
co-occurrence matrix values around a major
gray-level transition; (c) entropy, a measure of
the uniformity of matrix values; and (ci) comrela-
tion, a measure of linearity of the gray-level re-
bationship between rebated pixels. These four
features correspond to the smoothness or
coarseness as perceived by the sonologist.
In run-length histogram analysis, the number
of gray-level runs are counted by their length
and gray-level range. A run is a set of vertically
or horizontally neighboring pixels displaying
similar gray levels (Fig 6). We typically use run
lengths of 2 and 4 pixels. Parameters include
run percentages in the horizontal and vertical
directions, which characterize the distribution
of runs, and long run emphasis in the horizon-
a. b.
Figure 7. Fractal feature analysis. Histograni dis-
plays fractal dimensions of pixel intensities. The
height of each column is based on pixel intensity. A
perfect three-dimensional cube would he formed if
all pixel intensities were maximum. Instead, the
varying intensities ttll somewhere between a dimen-
sionality value of two and three.
1106 #{149}Imaging & Therapeutic Technology Volume 14 Number 5
Figure 8. Perfect fractals in nature. (a) Photograph of a seascape obtained at a distance of 3 in shows a sea fan
(Gorgonfa species) on the right with a branching pattern and irregular surfaces. If the sea fan were measured
with a 10-cm ruler, the overall perimeter would he much shorter than if it were measured with a I -cm ruler.
(b) Photograph obtained at a 0.3-m distance shows the branching patterns and irregular surfaces of a small por-
tion of the sea fan. These are similar to those seen when the whole colony is viewed (self-similarity). (Fig 8a and
8b courtesy of Steven R. Dent, MS. Department of Geology, University of Cincinnati, Ohio.)
tab and vertical directions, which describes the
frequency of occurrence of long runs.
Fractal feature analysis involves calculation
of the fractional dimension of the irregular sun-
faces created when the pixel intensities are dis-
played as a histogram (Fig 7). Benoit Mandel-
brot defmed fractabs as objects that are heteno-
geneous, self-similar, and impossible to measure
with a single unit of measure (33) (Fig 8). Al-
though perfect fractal surfaces have a constant
fractal dimension oven all ranges of scales, gray-
level images are imperfect fractal surfaces and
thus demonstrate a constant fractal dimension
over only those scales that reflect the size of
the anatomic structures of interest (34). The
software allows the scale to be changed if de-
sired.
A single textural feature or a combination of
them may be used to separate ROIs with me-
spect to change over time, pathologic change
versus normal tissue, on other variables. It is es-
sentiab to use repeatable machine settings and
to check machine calibration with a tissue-mim-
icking phantom. We currently hold the ma-
chine settings constant between subjects in lab-
oratory experiments and use the phantom data
for quality assurance. In the clinical setting, we
begin with standard machine settings, acquire a
phantom image, and then change the machine
settings to provide the best images possible for
each patient. Every time a machine setting is
changed, an additional phantom image is ob-
tamed. The phantom data are subjected to the
same textural analysis as the clinical data. Fea-
tunes with statistically significant differences in
the phantom data are discarded from further
>C
ci)
D%tile 10
0mean gray
I%tile 90
Tissue
September 1994 Allison ct al U RadioGraphics #{149}1107
Figure 9. (;raph displays fmrst-()rder
gray-level statistics for a population
of children undergoing intraopera-
tive neurosonography during nesec-
tion of mass lesions. The graph de-
picts the mean values for the mean
gray level, the 10th percentile, and
9()th I)erccntile of pixel intensities.
ABA =mass lesions other than tu-
mor, VG11 =normal gray matter,
NWM =normal white matter, PH =
calibration phantom, 1=tumor.
use. The phantom data may also be used to nor-
mabize the clinical data. Descriptive statistics
are followed by multiple analyses of variance,
regression analysis, or generation of receiver
operating characteristic curves. The results may
be displayed graphically (Fig 9) or as an image
overlay, similar to that used in color Doppler
imaging.
aCONCLUSION
Quantitative analyses. such as textural feature
analyses, allow additional digital manipulation
of the US B-mode data to overcome some of the
limitations of qualitative interpretation. These
techniques have already proved useful in inter-
pneting images of the eye, breast, kidney, and
liven. We are currently investigating these tech-
niques for analyzing images of the brain. Be-
cause certain patient conditions limit the use of
nonportable imaging modalities. we embrace
the use of off-line computers for acquiring addi-
tional information about iir critically ill pa-
tients.
.REFERENCES
1.Chivers RC. Tissue characterization. Ultra-
sound Med Biol 1981 7:1-20.
2. Insana MF, Garra BS, Rosenthal SJ. Hall 11.
Quantitative ultrasonography. Med Prog Tech-
nol 19S9; 15:141-153.
3. Insana MF, Wagner RF, (arra 135, Brown l)C,
Shawker 1H. Analysis of ultrasound image tex-
ture via generalized Rician statistics. Optical Eng
1986: 25:743-748.
4. Julesz B. Textons: the elements of texture per-
ception and their interactions. Nature 1981
290:91-97.
5. I)wyer SJ III, Stewart BK, Sayre jW. et al. Per-
formance characteristics and image fidelity of
gray-scale monitors. In: Honcyman Jc. Staah EV,
eds. Syllabus: a special course in computers for
clinical l)ractice and education in radiology. Oak
Brook. Ill: Radiological Society of North Amen-
ca, 1992: 117-124.
6. Insana MF, Hall Tj, Fishhack jL. Identifying
acoustic scattering sources in normal renal pa-
renchyma from the anisotnopy in acoustic prop-
erties. Ultrasund Med Biol 1991; 17:613-626.
7. Garra BS. Insana MF. Sesterhenn IA, et at.
Quantitative ultrasonic detection of parenchy-
mat structural change in diffuse renal disease. In-
vest Radiol 1994; 29:134-140.
8. (;hivers RC, Hill (;R. A spectral approach to ul-
trasonic scattering from human tissues: meth-
ods. objectives. and backscattenng measure-
1-net-It. Phys Med Biol I 975: 20:799-815.
9- Waag RC. Areview of tissue characterization
from ultrasonic scattenng. IEEE Trans Biomed
Eng 1 984: BME-3 1:884-893.
10. 1-laherkorn U, Zuna I, Lorenz A, et al. Echo-
graphic tissue characterization of diffuse paren-
chvnial liven disease: correlation of image struc-
tune with histology. Ultrason Imaging 1990:
12:155-170.
1 1. (;trra I3S, Insana MF. Shawker TH. Russell MA.
Quantitative estiniation of liver attenuation tnd
echogenicity: normal state versus diffuse liver
disease. Radiology 1987: 162:61-67.
I 2. (arrn BS, Insana MF. Shawker TH, et al. Quan-
titative ultrasonic detection and classification of
diffuse liver disease: Companson with human
observer pertorniance. Invest Radiol 1989;
24:196-203.
13. Carson P1.. Chiang EH, RuhinjM, Ct al. Pre- to
postnatal reduction in ultrasound attenuation co-
efficient of the liver. Invest Radiol 199 1 : 26:8-
I2.
14. Taylor KJW, Riely CA. Hammers L. et al. Quan-
titative US attenuation in normal liver and in pa-
tients with diffuse liver disease: inil)ortance of
fat. Radiology 1986; 16():65-7 1.
1 108 #{149}Imaging & Therapeutic Technology Volume 14 Number 5
15. Raeth U, Schlaps I), Limberg B, et al. Diag-
nostic accuracy of computerized B-scan texture
analysis and conventional ultrasonography in
diffuse parenchymal and malignant liver dis-
ease.JCtJ 1985: 13:87-99.
16. Chen CF, Robinson DE, Wilson LS, et al. Clini-
cal sound speed measurement in liven and
spleen in vivo. Ultrasn Imaging 1987; 9:22 1-
235.
17. Somnien FG, Stetson P, Chen HS. Prospects for
ultrasonic spectroscopy and spectral imaging of
abdominal tissues. J Ultrasound Med 1993;
12:83-89.
18. RubinJM, Carson PL, Meyer CR. Anisotnopic
ultrasonic hackseatten from the renal cortex. tJl-
trasound Med Biol 1988; 14:507-511.
19. (,arra BS, Krasner BH, Honii SC, et al. Impnov-
ing the distinction between benign and malig-
nant breast lesions: the value of sonographic
texture analysis. Ultrason Imaging I 993; 15:
267-285.
20. Kuhota M, Yamashita Y, lga M, Tajima T, Mi-
tomi I. In vitro estimation and imaging of at-
tenuation coefficients and instantaneous fre-
quency for breast tissue characterization. tlltra-
sound Med Biol 1988; 14:163-174.
21 .Vered Z, Mohn (A, Barzilai B, et at. US inte-
grated backscatter tissue characterization of re-
mote myocandial infarction in human subjects. J
Am Coll Cardiol 1989; 13:84-91.
22. OphirJ, Maklad NF, Bigelow RH. Ultrasonic at-
tenuation measurements of in vivo human mus-
etc. Ultrason Imaging 1982: 4:290-295.
23. Bergen 6, Laugien P. Pennin J. A field of appli-
cation for US attenuation: 1)uchenne muscular
dystropy. In: ThijssenJM, Mazzeo V, eds. Pro-
ceedings of the 5th European Community
Workshop on Ultrasonic Tissue Characterization
and Echographic Imaging. Nijmegen, The Neth-
enlands: University of Nijmegen, 1985; 77-83.
24. Romijn RL, Thijssen JM, Oostenveld BJ, Verbeek
AM. Ultrasonic differentiation of intraocular
melanomas: parameters and estimation meth-
ods. Ultrason Imaging 1991: 13:27-55.
25. ThijssenJM, Verbeek AM, Romijn RL, Dc Wolff-
Rousendaal D, Oostenhuis JA. Echographic dif-
fenentiation of histological types of intraocular
melanoma. Ultrasound Med Biol 1991 : 17:127-
138.
26. Lizzi FL, Feleppa EJ. In vivo ophthalmological
tissue characterization by scattering. In: Jhung
KK, Thieme GA, eds. Ultrasonic scattering in hi-
ological tissues. Boca Raton, Fla: CRC, 1993;
393-408.
27. OphirJ, Cespedes H, Ponnekanti Y, Li Y, Li X.
Elastography: a quantitative method for imaging
the elasticity of biological tissues. Ultrason Imag-
ing 1991; 13:111-134.
28. Weszka J, Dyer C, Rosenfeld A. A comparative
study of texture measures for terrain classifica-
tion. IEEE Trans Syst Man Cybern 1976; 6:269-
285.
29. Kruger RP, Thompson WB, Turner AF. Corn-
puten diagnosis of pneumoconiosis. IEEE Trans
Syst Man Cybern 1974; 4:40-49.
30. Pratt WK. Digital image processing, New York,
NY: Wiley, 1978; 323.
3 1 .Unser M. Sum and difference histogram for
texture classification. IEEE Trans Pattern Analy-
sis Machine Intell 1986; 18: 1 18.
32. Haralick RM, Shanmugam K, Dinstein I. Tex-
tural features for image classification. IEEE Trans
Syst Man Cybern 1973; 3:610-621.
33. Mandelbrot BB. The fractal geometry of na-
tune. New York, NY: Freeman, 1983; 14-19.
34. Chen CC, DaponteJS, Fox MD. Fractal feature
analysis and classification in medical imaging.
IEEE Trans Med Imaging 1989; 8:133-141.
