Quantitative measures of backscatter from human skeletal muscle: Changes with Duchenne's Muscular...

2
ABSTRACTS, ULTRASONIC IMAGING AND TISSUE CHARACTERIZATION SYMPOSIUM offline. Thus, objective image parameters can be combined with the subjective image analysis of the investigator. This combination of an objective and subjective image description is planned to achieve improved diagnosis. 1.3 APPLICATION OF NEURAL NETS TO ULTRASOUND TISSUECHARACTERIZATION, J. S. Ostrem, A. D. Valdes and P. D. Edmonds, Bioengineering Research and Electromagnetic Sciences Laboratories, SRIInternational, Menlo Park, CA 94025. Measurements of ultrasound speed andfrequencydependence of attenuation coefficient in human breast biopsy specimens were previously analyzed by two differentstatistical techniques. Thegoals were to discriminate normal from pathological tissues and benign from malignant tissue, if possible. 128 cases, comprising 50 normal,56 benign and 22 malignant wereanalyzed by classical discriminant analysis and classification and regression trees (CART). While sound speed was consistently identified as thevariable with the greatest power to discriminate normal from pathological tissues, considerable difficulty was found in discriminating benign from malignant tissues. Nevertheless, CART appeared morepromising thandiscriminant analysis. Another classification approach of interestis the neuralnetwork. Our data were presented to a network in a fully-connected, feed-forward configuration with 6 inputs (the variables), 2 hidden layers of 12 andeither6 or 9 units, respectively, and3 outputs (the categories). 75% of the data(comprising thetrainingset) were all correctly classified after adjustment of the network weights by a backpropagation algorithm that minimized mean-square classification error. Among thetestset,comprising the remaining 25% of thedata, 11112 normal,11/15 benign, and 4/4 malignant were classified correctly. 1112 normals was misclassified as benignand4/15 benigns were misclassified as normals. This performance was superior to that of the CART and discriminant analyses. This work was supported by PHS-NIH-NC1 grant CA34398 1.4 TISSUECHARACTERIZATION BY HIERARCHICAL CLUSTERING TECHNIQUES,N. H. Wang,J. T. Sheu, and B. Ho, Michigan State University, Ultrasound Research Laboratory, East Lansing, MI 48824. Most tissue characterization obtained from pulse-echo ultrasound is based on the acoustic impedance differenceat tissue interfaces. Unfortunately,it has beenshown that there is no significant variation of acoustic impedance between normal and cancerous tissues [11.Onthe otherhand, tumorand cancerous tumorhave quitedifferentattenuation characteristics and velocity propagation. However,only limitedsuccess has been foundby using these properties. Theobjective of this work isto employ more features to characterize biological tissues by theuse of hierarchical clustering methods. In our approach, five features from the echo return areextracted; they arethe total energy,central frequency,frequency at whichthe peak amplitude occurs, 3-dB bandwidth of the echo spectrum, and the correlation coefftcient between the incident andreflected signals. These features containnot only the information of acoustic impedance variation but also the attenuation and velocity characteristics. Based on these data, hierarchical clustering techniques such as singlelink, complete link and Ward’smethods [2] are used for cluster formation.Fromthe results of clustering, the typesof tissue canbe readily identified.A section of human brain with hemorrhaged tumor is used for the preliminary study. Color graphic is used to indicate the amount of feature variation. The indices of cluster validity will be reviewed. Techniques as well as experimental results will be presented. PI Birnholz, J.C. IEEE Ultrasonic Symposium, pp. 31-32(1972). PI Jain, A.K and Dubes, R.C., Algorithmsfor Clustering Data, Prentice-Hall, Inc., N. J. (1988). 1.5 QUANTITATIVE MEASURES OFBACKSCATTERFROM HUMAN SKELETAL MUSCLE: CHANGESWITH DUCHENNE’SMUSCULAR DYSTROPHY, Maria Helguera,’ JackG. Mottley,’ Shreedevi Pandya’ andRichard Moxley*, Rochester Center for Biomedical Ultrasound, ‘Department of Electrical Engineering, College of Engineering andApplied Sciences, and department of Neurology, School of Medicineand Dentistry, University of Rochester, Rochester, NY 14627. Preliminary studies haveshown that variations in ultrasonic properties such asbackscatter and attenuation of skeletal muscle maybe related to the physiologic state of those tissues, therefore holding greatpromise as indices for quantitative tissue characterization.This studywasdesigned to elucidate patterns of change in ultrasonic backscatter from skeletal muscle with disease. Five normal volunteers andfive patients suffering from Duchenne’s Muscular Dystrophy@MD) werestudied. A water-tilled 190

Transcript of Quantitative measures of backscatter from human skeletal muscle: Changes with Duchenne's Muscular...

ABSTRACTS, ULTRASONIC IMAGING AND TISSUE CHARACTERIZATION SYMPOSIUM

offline. Thus, objective image parameters can be combined with the subjective image analysis of the investigator. This combination of an objective and subjective image description is planned to achieve improved diagnosis.

1.3 APPLICATION OF NEURAL NETS TO ULTRASOUND TISSUE CHARACTERIZATION, J. S. Ostrem, A. D. Valdes and P. D. Edmonds, Bioengineering Research and Electromagnetic Sciences Laboratories, SRI International, Menlo Park, CA 94025.

Measurements of ultrasound speed and frequency dependence of attenuation coefficient in human breast biopsy specimens were previously analyzed by two different statistical techniques. The goals were to discriminate normal from pathological tissues and benign from malignant tissue, if possible. 128 cases, comprising 50 normal, 56 benign and 22 malignant were analyzed by classical discriminant analysis and classification and regression trees (CART). While sound speed was consistently identified as the variable with the greatest power to discriminate normal from pathological tissues, considerable difficulty was found in discriminating benign from malignant tissues. Nevertheless, CART appeared more promising than discriminant analysis.

Another classification approach of interest is the neural network. Our data were presented to a network in a fully-connected, feed-forward configuration with 6 inputs (the variables), 2 hidden layers of 12 and either 6 or 9 units, respectively, and 3 outputs (the categories). 75% of the data (comprising the training set) were all correctly classified after adjustment of the network weights by a backpropagation algorithm that minimized mean-square classification error. Among the test set, comprising the remaining 25% of the data, 11112 normal, 11/15 benign, and 4/4 malignant were classified correctly. 1112 normals was misclassified as benign and 4/15 benigns were misclassified as normals. This performance was superior to that of the CART and discriminant analyses.

This work was supported by PHS-NIH-NC1 grant CA34398

1.4 TISSUE CHARACTERIZATION BY HIERARCHICAL CLUSTERING TECHNIQUES, N. H. Wang, J. T. Sheu, and B. Ho, Michigan State University, Ultrasound Research Laboratory, East Lansing, MI 48824.

Most tissue characterization obtained from pulse-echo ultrasound is based on the acoustic impedance difference at tissue interfaces. Unfortunately, it has been shown that there is no significant variation of acoustic impedance between normal and cancerous tissues [ 11. On the other hand, tumor and cancerous tumor have quite different attenuation characteristics and velocity propagation. However, only limited success has been found by using these properties.

The objective of this work is to employ more features to characterize biological tissues by the use of hierarchical clustering methods. In our approach, five features from the echo return are extracted; they are the total energy, central frequency, frequency at which the peak amplitude occurs, 3-dB bandwidth of the echo spectrum, and the correlation coefftcient between the incident and reflected signals. These features contain not only the information of acoustic impedance variation but also the attenuation and velocity characteristics. Based on these data, hierarchical clustering techniques such as single link, complete link and Ward’s methods [2] are used for cluster formation. From the results of clustering, the types of tissue can be readily identified. A section of human brain with hemorrhaged tumor is used for the preliminary study. Color graphic is used to indicate the amount of feature variation. The indices of cluster validity will be reviewed. Techniques as well as experimental results will be presented. PI Birnholz, J.C. IEEE Ultrasonic Symposium, pp. 31-32 (1972). PI Jain, A.K and Dubes, R.C., Algorithmsfor Clustering Data, Prentice-Hall, Inc., N. J. (1988).

1.5 QUANTITATIVE MEASURES OF BACKSCATTER FROM HUMAN SKELETAL MUSCLE: CHANGES WITH DUCHENNE’S MUSCULAR DYSTROPHY, Maria Helguera,’ Jack G. Mottley,’ Shreedevi Pandya’ and Richard Moxley *, Rochester Center for Biomedical Ultrasound, ‘Department of Electrical Engineering, College of Engineering and Applied Sciences, and department of Neurology, School of Medicine and Dentistry, University of Rochester, Rochester, NY 14627.

Preliminary studies have shown that variations in ultrasonic properties such as backscatter and attenuation of skeletal muscle may be related to the physiologic state of those tissues, therefore holding great promise as indices for quantitative tissue characterization. This study was designed to elucidate patterns of change in ultrasonic backscatter from skeletal muscle with disease. Five normal volunteers and five patients suffering from Duchenne’s Muscular Dystrophy @MD) were studied. A water-tilled

190

ABSTRACTS, ULTRASONIC IMAGING AND TISSUE CHARACTERIZATION SYMPOSIUM

plastic standoff with ultrasonic transducer (10 cm focal length, broadband, 5 MHz center frequency) was applied to the anterior face of the upper arm over the belly of the biceps, perpendicular to the surface of the skin. The backscatter spectrum from a 3.75 mm portion of muscle about 1.5 cm deep from the skin was recorded at different sites. (Freq. range: 2-8 MHz in 1 MHz steps). The spectra were reduced to decibels relative to a planar, stainless steel reflector and the frequency average, or integrated, backscatter was calculated. Average integrated backscatter was obtained by averaging at each site, then all sites within a subject, then over all subjects. The value obtained for normal volunteers was -56.7 + 1.7 dB (mean &- sem, n = 5), while the level obtained in patients with DMD was -40.0 + 2.8 dB (mean f sem, n = 5). Integrated backscatter from afflicted patients was significantly higher than that of normal volunteers, with an average difference of 16.7 + 3.28 dB (mean + sd, p < 0.01). The ultrasonic method reported in this work may prove to be very useful as a new application of a noninvasive technique in the diagnosis and monitoring of muscle diseases in substitution of current invasive techniques.

1.6 LIVER BACKSCATTER: A FRACTAL MODEL, T.A. Tutbill and K.J. Parker, Rochester Center for Biomedical Ultrasound, University of Rochester, Rochester, NY 14627.

Backscatter from the liver has been attributed to a variety of structures including collagen, portal triads, hepatocytes, fat globules, and other sized “scatterers” and small geometries. Some researchers have approximated these scattering structures as small, uniform scatterers in a semiregular architecture. However, a realistic assessment of the liver shows that the increased impedance mismatch should correspond to the collagen-sheathed, parallel branches of the portal vein and hepatic artery, which terminate in the portal triads. Plastic vascular casts of the liver show an arborization of blood vessels with branches of various sizes.

In our analysis, the vasculature is modeled as a fractal structure of discrete branching scales with an exponential probability density function. Assuming a fractal dimension consistent with physiological data, a Gaussian autocorrelation for each branch size, and a linear superpositioning of spectra, the resulting backscatter from the fractal vascular tree displays a frequency dependence with a power law factor ranging between 1 and 2, depending on branching features. The spectral slope shows good agreement with experimental data using calf and rabbit livers, with and without enhancement of the portal vein and hepatic artery branches. This work illustrates the potential for tissue characterization based on backscatter from a fractal vasculature.

1.7 IN VIVO ULTRASONIC MEASUREMENT OF MICROSCOPIC ANATOMICAL CHANGES IN OBSTRUCTED KIDNEYS, M.F. Insana,’ T.J. Hall,’ J.G. Wood2 and L.Y. Cheung2, Departments of *Radiology and ‘Surgery, University of Kansas Medical Center, Kansas City, KS 66103.

Properties derived from acoustic backscatter measurements are examined for dog kidneys undergoing changes in hydrostatic pressures induced intraoperatively. These experiments are a part of our investigations into the sources of acoustic scattering in normal renal tissues. The objective of the study was to verify in vitro experimental results reported at this meeting last year, which suggest that glomeruli are the principal scatterers at ultrasonic frequencies below 5 MHz and renal tubules and/or blood vessels above 5 MHz.

Following unrelated experiments, normally-perfused dog kidneys were exposed and scanned, in situ, to measure the average scatterer size and several other acoustic properties at frequencies below and above 5 MHz. Diuresis was induced using marmitol prior to complete temporary obstruction of the left ureter. Arterial pressure, ureteral pressure, and urine flow was measured along with the acoustic measurements, throughout the experiment. Unilateral ureteral obstruction was maintained for 10 minutes before being released for 5 minutes; this pattern was repeated several times. Ureteral obstruction increased the pressure in the ureter from less than 10 mm Hg to values between 70 and 120 mm Hg, depending on the animal. We found that changes in the scatterer size estimates for frequencies below 5 MHz varied in proportion to the pressure in the ureter, where the magnitude of the change was between 12 and 30%. Subsequently, we found that the acoustically-determined scatterer size corresponded to the average glomerular diameter, further supporting the theory that glomeruli are the principal scattering structure below 5 MHz. At frequencies greater than 5 MHz, very little change in scatterer size was observed, suggesting that arterioles, rather than tubules, may be the principal scatterers at high diagnostic frequencies. Experiments using vasodilators and vasoconstrictors are expected to further delineate the contributions of renal blood vessels and tubules in ultrasonic backscatter.

This work was supported in part by the Whitaker Foundation.

191