IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. BME-33, NO. 7, JULY 1986

Quantitative Tissue Characterzation Based on Pulsed-Echo Ultrasound Scans

EUGENE WALACH, C. N. LIU, SENIOR MEMBER, IEEE, ROBERT C. WAAG, SENIOR MEMBER, IEEE, AND

KEVIN J. PARKER, MEMBER, IEEE

Abstract-This paper describes a novel technique for estimating ul-trasonic attenuation coefficients. The technique first employs a histo-gram analysis to estimate the number of tissues present and then uti-lizes a maximum likelihood criterion to assign attenuation values, thusproducing an image of attenuation. Simulated B-scan data and clinicalB-scan data are used to illustrate the method. The results show thatimages representing an intrinsic tissue parameter can be producedwhen the basic model is valid.

I. INTRODUCTIONU LTRASONIC imaging has already had a major im-

pact on diagnosis in medicine. However, currentclinical imaging techniques and computer image enhance-ments produce only qualitative images which require ex-perience for interpretation [1]. The purpose of this paperis to describe a new method for estimating the tissue at-tenuation coefficient based entirely on conventional ultra-sonic B-scan measurements [2].We will assume a simple exponential model for the

backscatter echo amplitude from the ith pulse of a sectorscan [3]-[6]:

Ei(x) = Xu1(x) exp L-2 aci(r) drj(1)where a represents the scattering coefficient, x representsthe (nonzero) distance to the transducer, ae stands for theattenuation coefficient, and E0 is a constant of proportion-ality. The factor of two in the exponent in (1) arises fromthe roundtrip to the reflector and back to the transducer.In our model, transducer beam width and pulse length ef-fects, which produce amplitude fluctuations sometimescalled speckle or texture [7], [8], are represented by theassumption that the backscatter echo amplitude is a ran-dom variable.By sampling both sides of (1), the measured data can

Manuscript received June 6, 1985. This work was supported in part byNSF Grants ECS 80 17683 and ECS 84 14315, NIH Grant CA 39516,NATO Grant 93 81, and the Diagnostic Ultrasound Research LaboratoryIndustrial Associates.

E. Walach was with the IBM T. J. Watson Research Center, YorktownHeights, NY 10598. He is now with the IBM Haifa Scientific Center, Haifa,Israel.

C. N. Liu is with the IBM T. J. Watson Research Center, YorktownHeights, NY 10598.

R. C. Waag and K. J. Parker are with the Departments of ElectricalEngineering and Diagnostic Radiology, University of Rochester, Roches-ter, NY 14642.

IEEE Log Number 8608288.

be represented as a system of nonlinear equations-oneequation for every time sample, which in turn correspondsto a certain picture element (pel):

Eij =Eo ij exp L-2 E aUl (2)

where Eij is the amplitude of the echo received from thepel ij (the ith interrogation of the jth sample volume), jand aij are backscattering and attenuation coefficients, re-spectively, and the summation is performed along the pathof the ultrasonic ray. By simply taking logarithms of bothsides, one can rewrite (2) as a system of linear equations

ln E* = ln Eo0 - 2Aa (3)

where E* stands for the vector of measured echo samples(multiplied by corresponding distances to the transducer),a and a are vectors of backscattering and attenuation coef-ficients, and A is a matrix (with entries 0 or 1) determinedby the geometry of the problem. A B-scan image providesthe values for the left side of (3). The matrix A is knownfrom the experimental setup. The goal is to find a and a.Assume for convenience and without loss of generality

that the image is square of size N x N. Hence, the systemof equations (3) will have 2N2 unknowns (backscatter andattenuation coefficients for every pel of the image) andpN2 equations where p is the number of different projec-tions of the area under consideration. Naturally, the ex-istence of a unique solution to (3) is contingent on theproblem being nonsingular. This means that p has to beat least not smaller than two [3]. Unfortunately, in mostpractical cases, even a relatively large p does not ensurea reasonable estimation of af and a. Indeed, in most prac-tical cases, the eigenvalue spread associated with the sys-tem of equations (3) is extremely high. Moreover, the sig-nal-to-noise ratio of the ultrasonic data is quite low [7],[8]. Consequently, large computational artifacts may dis-tort an image of attenuation and backscatter coefficient[3]. In order to resolve this issue, one has to redefine themathematical formulation of the problem so that the num-ber of degrees of freedom will be reduced. Fortunately,since the biological object to be imaged may consist of arelatively small number of uniform tissues, we can oftenfind regions of constant a and then estimate the corre-sponding a.

In this paper, we will first describe a way to determine

0018-9294/86/0700-0637$01 .00 © 1986 IEEE

637

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. BME-33, NO. 7, JULY 1986

the number of different tissues present in any given im-age. Then we develop a way to partition the image into anumber of mutually exclusive areas-one area for everytissue present. Once such a partition is achieved, we pro-ceed to compute an attenuation coefficient for every tissuearea using a maximum likelihood algorithm. Finally, threeexamples are presented in order to illustrate the perfor-mance of the approach.

IL. STATISTICAL CHARACTER OF THE ULTRASONIC ECHOAMPLITUDES

Consider first the distribution of the echo amplitudesreceived from a statistically homogeneous region of ran-domly positioned scatterers. As already obseryed, despitethe homogeneity of the region of interest, one can expectamplitude fluctuations which arise from within the reso-lution cell formed by the beam width and pulse length asthe transducer is scanned through the region of interest.When the number of scatterers within one resolution ele-ment (pel) is large and the phases of the scattered wavesare uniformly distributed between 0 and 2-7-, the ultrasonicecho amplitude, as noted by previous investigators [7],[8], will have a Rayleigh distribution:

V -V2/2sp(v)=-e v.>-0

-0 v < 0 (4)

where v stands for the echo amplitude, p is the probabilitydistribution, and s is a constant called the Rayleigh coef-ficient which can be expressed in terms of the mean bythe relation E[jv] = (7rsI2)112.Next consider a more realistic case in which the object

consists of a number of different tissues. Each tissue willhave different backscattering properties, i.e., differentRayleigh coefficients. Hence, the overall probability dis-tribution will consist of a number of superimposed Ray-leigh distributions.The first question is how many different tissues are pres-

ent in any given image and what are the correspondingRayleigh coefficients. An experienced physician mighthave an immediate answer to this question based on theknowledge of the anatomical structure of the imaged tar-get and accumulated statistics of the previous studies.However, in this investigation, we will proceed under theassumption that no such a priori knowledge is available.Hence, this issue must be resolved by the analysis of theimage itself.The Rayleigh probability distribution function has a

single peak located exactly at the square root of the cor-responding Rayleigh coefficient. The histogram of the im-age of a nonhomogeneous object will have a number ofpeaks. The number of different tissues present will beequal to the number of such peaks and each peak impliesa Rayleigh coefficient.

This observation suggests a simple procedure for find-ing the Rayleigh coefficients of the various tissues presentin a given image from histograms. The preparation of ahistogram requires a choice of a single arbitrary parame-

ter: "the interval" or "cell size" [9]. Following [9], wechoose the cell width as twice the interquartile range ofdata divided by the cube root of the sample size. In ourexperience, this rule proved to be very efficient and use-ful.Once the number and character (Rayleigh coefficients)

of all the tissues present have been established, it is pos-sible to decide to which tissue every pel belongs. As adecision rule, we have adopted the principle of maximumlikelihood which is known to have certain optimalityqualities [10].Assume that we have established the existence of the K

different tissues with the corresponding Rayleigh coeffi-cients being S1, * - *, SK. We wish to "assign" the pelij which resulted in the echo Eij to one of the possibletissues. Thus, we compute the probabilities

PIc = -E -E2/2SkSk

K 2 k 2 1. (5)

Each number Pk provides the probability of obtaining theecho Eij from the pel belonging to the tissue number k.The maximum likelihood principle assigns the pel ij to thetissue k which has the highest probability:

Pk= max {P, - *, PK}. (6)Introducing a spatial filtering to reduce the error of as-

signing the pels, we obtain the following algorithm.1) Choose a window Wij which will include all the pels

which are relevant to the tissue assignment of the pel ij.2) Assign weight tr to every pel r inside this window.3) For every K > k 2 1 and every pel in the window,

compute the probability Prk of obtaining echo Er if pel rbelonged to the tissue k.

4) Compute

Pk= H tr

Wii(7)

5) Use the maximum likelihood principle, i.e., assignpel ij to the tissue which resulted in the maximum prob-ability Pk-

In our simulations, we have chosen small windows of3 x 3 or 5 x 5 pels (which correspond roughly to theresolution which is about 1 mm in the axial direction alongthe wave and 3 mm in the lateral dimension) and assignedthe uniform weight (tr = 1) to every pel inside the win-dow.

III. ESTIMATION OF THE TISSUE ATTENUATIONCOEFFICIENT

For practical reasons, let us assume that the tissues areindependent from each other and that the data can be rep-resented in the form depicted in Fig. 1. The tissue underconsideration is crossed by a number P of ultrasonic rays.Each ray i, P 2 i 2 1, contributes a block of ki pelswhich are located inside the tissue. We will denote theechos of the block number i as Eil, Ei2, * * *, Eik,. Sincewe do not know what was the impact of the path passedbefore each ray arrived at the area of interest, we have to

638

WALACH et al.: TISSUE CHARACTERIZATION BASED ON ULTRASOUND SCANS

El, Ei1l PE i

-Block 1 ... Block Block P

Elk Eiki EPkp

y

Fig. 1. Schematic structure of the data.

assume that each block may start at some arbitrary levelEi. However, from that point on, all the echos from allthe blocks are assumed to follow an exponential depen-dence on the same attenuation coefficient oa. Hence, forevery tissue, we will have, to begin with, E.f 1 ki datasamples (equations) in P + 1 unknowns. Our goal is tofind an attenuation coefficient which will minimize a cer-

tain cost function.Assume for the moment that we know the value of the

attenuation coefficient a and a known accumulated atten-uation Ai for every block i. Then compensate each echosample for the attenuation effect. The compensated echofrom the pel ij will have the value

AiE0 e2(j- )a. (8)

Finally, equating to zero the derivative of (11) relative tothe a and substituting (12) yields the "optimality" con-dition

ki

P P kiE EJ2(j - 1) e4(i-l)"0.5 E ki(ki - 1) = E i= i1i- ~ ~~~~~~~~Ei-ge4( - I)a

j = I

(13)

Equation (13) is nonlinear in ca. In order to obtain a closedexpression for az, we will assume that the attenuation coef-ficient is small (relative to the unity), and therefore all theexponentials in (13) can be approximated by the first twoterms in the Taylor series expansion. Hence, (13) yields

p

ENOii-1 (14)

ZDNii= 1

where

NOi = 0.5k (ki- 1)-ki-

DNi = 4ki - 2a

(15)

(16)The probability of obtaining this value from the tissue withthe Rayleigh coefficient s is

AiE e2(j1)e A2Eij2 e4' -l)iPij = exp - 2s j (9)

Assuming that all the data samples are independent of eachother, the overall probability of obtaining the existing setof measurements is

P ki

c II Pe* (10)i=1 j=1

The maximum likelihood principle' requires the choice of

a and Ai,, P > i > 1, such that probability (10) will bemaximized. Maximizing C is equivalent to the maximi-zation of

P ki

lnC= E lnpiji=l j-l

P k,

=ln C0+ ZlnAi

+ 2(j- i)a-I

" e4(i ')a> (12s

where Co is a constant independent of Ai, P >- i >- 1, and

a. Equating to zero the derivative of (11) relative to Aiyields for every P > i - 1

A?s

kiki

E E2 e4(j- I)aj=l

Ki

ai= E Etj=lKi

bi= (j -1)Ej=1

Ki

ci= E (j - 1)2Ei.j=l

(17)

Utilizing (14)-(17), the estimated value of the tissue at-tenuation coefficient can easily be computed. Moreover,since the bulk of computations is performed indepen-dently for every block of data, our algorithm can be easilyimplemented on a parallel processor for maximum speed.The accuracy of this algorithm is analyzed in the Appen-dix.

IV. EXAMPLES AND DISCUSSIONA B-scan image model which consists of two oval cysts

on uniform background is shown in Fig. 2. The Rayleighcoefficient of the right-hand cyst was chosen to be one.

The choice for the left-hand cyst and for the backgroundwas 10 000 and 100, respectively. Attenuation coeffi-cients were chosen to be 0, 0.01, and 0.0025 neper/pel,respectively. The B-scan image was created by drawingindependent amplitude samples from the pools of randomvalues having Rayleigh distributions with the chosen coef-ficients. The attenuation effects were simulated by mul-tiplying each amplitude value by the integral of the atten-uation along the ray path. The procedure described in thelast section was used to estimate the tissue attenuation

and

639

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. BME-33, NO. 7, JULY 1986

(a) *(b)Fig. 2. Simulated B scan (b) of an object (a).

TABLE IESTIMATED ATTENUATION COEFFICIENT OF THE SIMULATED TARGET

AreaAttenuationCoefficient Right Cyst Left Cyst Background

True Value 0 0.01 0.0025Estimation 0.0003 0.0095 0.0026Error 0.00028 0.0008 0.0002

coefficients for this simulated B-scan image. The resultsare presented in the Table I. In this case, clearly, the es-timated values (the second row) constitute an excellentapproximation of the actual values (the first row). Thethird row of Table I presents our estimation of an error(the square root of the variance). Note that the lowest er-ror can be expected for the background which has thelargest area. On the other hand, the right-hand cyst has asmaller area and has a correspondingly larger estimationerror. However, since the scan direction is from left toright, this cyst includes long data blocks. Hence, esti-mated error remains relatively low. Finally, the left-handcyst has the worst error which reflects small area and rel-atively short data blocks. In all three cases, the actual es-timation error was well within reasonable expectations.

Next, the procedure was applied to B scans of humanliver. The B scans were obtained using an Octoson im-aging instrument in a way previously described [2], [1 1].Basically, backscattered RF waveforms were digitized andstored along with beam position information and time-varying gain settings. During subsequent digital process-ing, the waveforms were envelope detected and then theeffects of time-varying gain and amplifier compressionwere removed. No correction for beam diffraction wasemployed because the effect of diffraction was known tovary amplitudes less than 10 percent throughout the 6 cminterval surrounding the transducer focus which was sit-uated centrally in the livers being studied.A B-scan image of a normal liver is shown in Fig. 3.

After the waveforms used to construct this image weredecompressed, a histogram of the amplitudes was ob-tained. An analysis of this histogram indicated that sixdifferent regions of tissue were present. Our computationof attenuation for each region produced the image pre-

Fig. 3. Digitized clinical B scan of a normal liver.

Fig. 4. Attenuation imaging computed from echos in Fig. 3.

sented in Fig. 4. In this image, a linear gray scale is usedto represent attenuation in decibels with black equal to 0dB/cm and white equal to 4.5 dB/cm. High attenuation(about 1.75 dB/cm) was found near the edges of the liver,while the interior of the liver is relatively homogeneouswith attenuation in the range of 0.6-0.8 dB/cm. The val-ues in the interior are comparable to a value of 1.0 dB/cm found at 2 MHz by a spectral decay method treatedelsewhere [2], [11] and the high attenuation values sur-rounding the liver are attributed to boundary effects.

In Fig. 5 is a B scan of a liver known to have highultrasonic attenuation. Waveform decompression, ampli-tude histogram analysis, and partitioning into six differenttissue regions for this study produced the image of atten-uation given in Fig. 6 for which the gray scale assign-ments are the same as those in Fig. 4. The attenuation for

640

WALACH et al.: TISSUE CHARACTERIZATION BASED ON ULTRASOUND SCANS

Fig. 5. Digitized clinical B scan with high attenuation.

tissue attenuation coefficient from the expressions (14)-(17). Our aim is to establish a simple approximation ofthe error of estimation based on a single block of K datasamples TI, T2e2A , TKe 2(K -)cwhere Ti, T2,TK are mutually independent random variables obeyingthe Rayleigh distribution with the Rayleigh coefficientequal to a certain positive number s. In order to achievethis result, certain simplifying assumptions and approxi-mations are required. However, despite this heuristic na-ture of our derivation, there is an excellent agreement be-tween the theory and the simulation results.

Utilizing expressions (14)-(17) for a single block of Kdata samples yields the following expression for the es-timated attenuation coefficient & as a function of the ac-tual attenuation coefficient a:

(K - 1) a(a) - 2b(u)

8 (c(a) - (a))

whereK

a(a) = E T2 e4(J)j=I

Kb(a) = E TJ(j - 1) e-4(j- )

1=1

Kc(a) = E Tj2(j- 1)2 e-4(-1).

j=1(A.2)

Expanding (A. 1) in the Taylor series around ax = 0 yields

& =fO +f1a +f2U2 + -- (A.3)For small attenuation coefficient a, the error resultingfrom the high-order terms is negligible. Hence, we willconcentrate only on the error introduced by the first biastermfo. By examining (A. 1), we can express the bias termas

Nofo=-Fig. 6. Attenuation imaging computed from Fig. 5. (A.4)

a broad central area below the abdominal wall is 1.3 dB/cm, which is the same as that given for the region at 2MHz by a spectral decay method.Banding of attenuation with range is evident to different

degrees in Figs. 4 and 6. This banding is not thought tobe a consequence of liver structure. Rather, the bandingis thought to arise from a combination of factors whichinclude beam diffraction, imperfect amplifier decompres-sion, and poor signal-to-noise ratio in posterior regions.While the specific influence of various factors merits

additional investigation, our results illustrate a new waythat quantitative ultrasonic images may be obtained fromconventional pulsed-echo measurements.

APPENDIX

The purpose of this Appendix is to evaluate the esti-mation error which can be expected while computing the

where

No = (K - 1) a(0) - 2b(0)

/0 _ ___

D= 8(C(O) -a(0)) (A.5)

Consider individually the behavior of the numerator Noand the denominator Do. For the large values of K, thedenominator will be also large and we have

[K KK

E[a(O)IJ = E Y, T,2j = Eli2 Y, 1 = 2Ks (A.6)

E[b(O)] = E -,j 1)] = E[Tj2]j, (j- 1)1j= I I j=1 - 1

641

(A.7)= K(K - I)s

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. BME-33, NO. 7, JULY 1986

-K -KE[c(O)] = E E2Tj(j - 1)2 = E[Tj2] K (j-1)2

-i=l j=l~~~ J)

= 3 K(K - 1)(2K - 1)3

K K KE[b2(0)] = T4(j _ 1)2 + E T2Eb =E[Z)4j )2

j= i=l

- (E[TJ4] E (j - 1)2 + E[T;])2J = 1

(A.8)

Tr2(i- 1)]

0b-

e

b- 10-4 -w

0*

(I.C20f0_

20

S

theoretical

computerestimation

SS

0

40Samples in Block

100

Fig. 7. Dependence of estimation error on the length of data block.

K

(j - 1) Z (ii*j

Substitution of (A.9), (A.13), and (A.14) into (A.12)yields

- 1))

= - K(K - 1)(3K2+ K - 2).3 (A.9)

Substituting into (A.4) the corresponding expected valuesinstead of a(0), b2(0), and c(0), we can express the de-nominator as

DO - 8( K(K - 1)(2K - 1) - K(K - 1)6

(A. 10)

On the other hand, for small values of K, the denomi-nator will be very small. Thus, according to (A.6) and(A.7), we have

E[No] = (K - 1) E[a(0)] - 2E[b(0)] = 0. (A. I 1)However, the numerator will have a nonzero variance:

E[N0] = (K - 1)2 E[a2(0)] - 4(K - 1)E[a(0) b(0)] + 4E[b2(0)]. (A. 12)

In order to evaluate the expression (A. 12), we noteK K K

E(a2(0)] = ZE T4 + Z T2j=l j-1 i=l

i *j

= 8s2K + 4s2K(K - 1) = 4s2K(K + 1) (A.13)

and

K K

Z T2 ETZ2(i - 1)j-1 i-i=I

1*1/ K

= (E[J4] E (j-_ 1)) + (E[T21)2K KZE Z (i - 1)(j - 1)j=l i-l

i*j

= s2(4K(K - 1) + 2K(K - 1)2)= 2s2K(K2 - 1).

N2] 423K2 3EIN6] = -~(K-I) K(K+ 1) -K.

Hence, we conclude that for a small attenuation coeffi-cient a and for a large block length K, the bias term fowill have a small mean value, i.e.,

(A. 16)fo EN0] = 0E[D0]

and a nonzero variance which can be approximated as

E ] =-E[ND2] - (E [NO2] -43-E[D6] -(E [DO]) - 4K3

In order to verify the expression (A. 17), a computersimulation was performed. The attenuation coefficient wascomputed utilizing expression (A. 1) for blocks of datavarying in length from 25 to 100 samples. The error be-tween the actual and estimated attenuation coefficients wascomputed. The process was repeated 100 times, each timewith. a totally independent data set. Then the mean-squareerror was computed as a function of the block length. Theresult is presented in Fig. 7. In the log-log scale, theexpression (A. 17) takes the form of the straight line. Thecircles represent the mean-square error measured from thesimulation. Despite some crude approximations which hadbeen adopted in our derivations, an excellent agreementbetween the theoretical and experimental results can beobserved.

ACKNOWLEDGMENTThe authors acknowledge help from Dr. R. Gramiak

and Dr. R. M. Lemer for useful discussions of the results.

REFERENCES

[1] C. N. Liu, M. Fatemi, and R. C. Waag, "Digital processing for im-provement of ultrasonic abdominal images," IEEE Trans. Med. Im-aging, vol. MI-2, pp. 66-75, June 1983.

[2] K. J. Parker and R. C. Waag, "Measurement of ultrasonic attenua-tion within regions selected from B-scan images," IEEE Trans.Biomed. Eng., vol. BME-30, pp. 431-437, Aug. 1983.

[3] E. J. Farrell, "Backscatter and attenuation imaging from ultrasonicscanning in medicine," IBM J., vol. 26, pp. 746-758, Nov. 1982.

[4] F. A. Duck and C. R. Hill, "Acoustic attenuation reconstruction from(A. 14) backscattered ultrasound," in Computer Aided Tomography and Ul-

/ K

.1 zE1(A.15)

* (3K2 + K- 2)) = K3.(A.17)

E[a(O) b(O)] = E L_E T14]

642

WALACH et at.: TISSUE CHARACTERIZATION BASED ON ULTRASOUND SCANS

trasonics in Medicine, J. Raviv, J. F. Greenleaf, and G. T. Hernman,Eds. Amsterdam: North-Holland, 1979.

[5] R. W. Rowe, "Ultrasonic image reconstruction," IBM U.K. Sci.Cen., Winchester, England, Rep. UKSC 107, 1981.

[61 J. F. Havlice and J. C. Taenzer, "Medical ultrasonic imaging: Anoverview of principles and instrumentation," Proc. IEEE, vol. 67,pp. 620-641, Apr. 1979.

[7] C. B. Burckhardt, "Speckle in ultrasound B-mode scans," IEEETrans. Son. Ultrason., vol. SU-25, pp. 1-6, Jan. 1978.

[8] R. F. Wagner, S. W. Smith, and H. Lopez, "Statistics of speckle inultrasound B-scans," IEEE Trans. Son. Ultrason., vol. SU-30, pp.156-163, May 1983.

[9] D. Freedman and P. Diaconis, "On the histogram as a density esti-mator: L2 theory," Z. Wahrscheinlichkeitstheorie verw. Gebiete, vol.57, pp. 453-476, 1981.

[10] G. C. Goodwin and R. L. Payne, Dynamic System Identification: Ex-periment Design and Data Analysis. New York: Academic, 1977.

[11] K. J. Parker, R. M. Lemer, and R. C. Waag, "Attenuation of ultra-sound: Magnitude and frequency dependence for tissue characteriza-tion," Radiology, vol. 153, pp. 785-788, Dec. 1984.

Eugene Walach was born in Lvov, U.S.S.R., onFebruary 2, 1952. He received the B.Sc., M.Sc.,and D.Sc. degrees in electrical engineering, fromthe Technion-Israel Institute of Technology,Haifa, in 1973, 1975, and 1981, respectively.

During the years 1981-1983 he was with theInformation System Laboratory, Stanford Univer-sity, as a Chaim Weizmann Postdoctoral Fellow.In 1983-1984 he was a Visiting Scientist at theIBM T. J. Watson Research Center, YorktownHeights, NY. He is currently a staff member of

the IBM Israel Scientific Center, Haifa, Israel. His research interests arein the areas of image and signal processing, adaptive systems, and analysisof multivariable systems.

Dr. Walach is a member of Sigma Xi.

0 S_, C. N. Liu (M'67-SM'83) received the Ph.D. de-~ ggree in electrical engineering from the University

of Illinois, Chicago, in 1961.He is currently a Research Staff Member and

the Manager of Image Construction and Presen-tation Systems in the Computer Sciences Depart-ment of the IBM T. J. Watson Research Center,Yorktown Heights, NY, responsible for the re-search and advanced development of image/graphics systems for engineering/scientific appli-cations. From 1978 to 1981 he was Manager of

Medical Ultrasound, engaging in the development of computer techniquesfor extracting quantitative diagnostic information from pulsed-echo ultra-sound scans. From 1961 to 1977 he worked on a number of projects in-volving the development of pattern recognition and image processing tech-

niques and systems. From 1968 to 1969 he was on leave from IBM Researrhas a Visiting Associate Professor at the School of Electrical Engineering,Purdue University, Lafayette, IN. His current research interests includeengineering and scientific image/graphics workstations, image processing,and pattern recognition systems.

Dr. Liu is a member of the Association for Computing Machinery.

J Robert C. Waag (S'59-M'66-SM'83) was bornin Upper Darby, PA, in 1938. He received theB.E.E., M.S., and Ph.D. degrees from CornellUniversity, Ithaca, NY, in 1961, 1963, and 1965,respectively.

After completing his Ph.D. studies, he becamea member of the Technical Staff at the Sandia Lab-oratories, Albuquerque, NM, and then served inthe United States Air Force from 1966 to 1969 atthe Rome Development Center, Griffiss Air ForceBase, NY. In 1969 he joined the faculty at the

University of Rochester, Rochester, NY, where he is now a Professor inthe Department of Electrical Engineering, College of Engineering and Ap-plied Science, and also holds an appointment in the Department of Ra-diology, School of Medicine and Dentistry. His recent research has dealtwith computer-based processing of ultrasonic signals and the use of ultra-sonic scattering for determination of tissue characteristics.

Dr. Waag is a Fellow of the American Institute of Ultrasound in Med-icine and a member of the Acoustical Society of America and the Associ-ation for Computing Machinery.

Kevin J. Parker (S'79-M'81) was born in Roch-.s_ ester, NY, in 1954. He received the B.S. degree

in engineering science, summa cum laude, fromS.U.N.Y., Buffalo, in 1976, and the M.S. andPh.D. degrees in electrical engineering andbiomedical ultrasonics from M.I.T., Cambridge,in 1978 and 1981, respectively.

From 1981 to 1985 he was an Assistant Pro-fessor of Electrical Engineering at the Universityof Rochester; currently he holds the title of As-sociate Professor. His research interests are in ul-

trasonic tissue characterization, medical imaging, and general linear andnonlinear acoustics.

Dr. Parker was the recipient of a National Institute of General MedicalSciences Biomedical Engineering Fellowship (1979), Lilly Teaching Fel-lowship (1982), and Whitaker Foundation Biomedical Engineering GrantAward (1983). He serves as Chairman of the Rochester Section of the IEEEEngineering in Medicine and Biology Society, a member of the IEEE Son-ics and Ultrasonics Symposium Technical Committee, and as reviewer andconsultant for a number of journals and institutions. He is also a memberof the Acoustical Society of America.

643