Ultrasonic tissue characterization of breast tumors using pattern recognition techniques

Ultrasonic tissue characterization of breast tumors using pattern recognition techniques

ABSTRACTS, ULTRASONIC ItWGING AND TISSUE CHARACTERIZATION SYMPOSIUM classification rates of 72 to 80 percent for the canine and 56 to 63 percent fo...

207KB Sizes 0 Downloads 5 Views

Recommend Documents

No documents
ABSTRACTS,

ULTRASONIC ItWGING AND TISSUE CHARACTERIZATION

SYMPOSIUM

classification rates of 72 to 80 percent for the canine and 56 to 63 percent for the human, but required more features to achieve these rates. Application of the nearest-neighbor rule resulted in higher rates of In the case of the correct classification for all the transforms used. canine samples, the identity transform performed best, achieving 91 percent correct classification with only 12 of an original 1024 features retained. For the human samples the DCT performed best with an 87 percent correctclassification rate and six of an original 1024 features retained. Research was supported in part by NASA contract NAS g-15452 and NIH Biomedical Sciences Support grant RR07114. THE EARLY DETECTION OF BREAST CANCER: AN ULTRASONIC APPROACH USING RF WAVEFORM ANALYSIS VIA PATTERN RECOGNITION, Morris S. GoodI, Joseph L. Rosel, and Barry B. Goldberg*, lDrexe1 University, Philadelphia, PA 19104 and *Thomas Jefferson University Hospital, Philadelphia, PA 19107. The purpose of this research effort is to define an accurate noninvasive technique for the diagnosis of malignant breast disease. Fluid areas were not of concern directly since present techniques were capable of Data collected for this distinguishing these areas by B-scan imaging. study, therefore, emphasize classification of solid tissue areas as either benign or malignant. Pattern recognition techniques such as the Fisher linear discriminant were used to determine various algorithms for classifying a tissue area. Tissue features used to develop the discriminant function were derived from the rf signal and evaluated by its Probability Distribution Function (PDF) for its decisive nature. Only features having a physical explanation to These its distribution and/or high discriminant value were retained. features were then used to develop an algorithm according to the pattern recognition method being implemented. Resulting algorithms were tested by the Jack-Knife technique to determine unbiased performance indices. Results from a 100 tissue area population provided sensitivity and specificity values of 95 percent and 67 percent respectively. ULTRASONIC TISSUE CHARACTERIZATION OF BREAST TUMORS Finettel, Alan RECOGNITION TECHNIQUES, Steven I. William Swindelll '* and Kai Haberl, lDepartment Arizona Health Sciences'Center and *Optical Sciences Center, Arizona, Tucson, AZ 85724

USING PATTERN R. Bleier*, of Radiology, University of

We consider the classification of in vivo human breast tumors as a problem in statistical pattern recognition. *% goal is the determination of accurate classification rules and confidence intervals for quantitative computer diagnosis of breast cancer using a conventional ultrasound imaging system. The basic approach is to measure properties of the (raw) backscattered RF waveform, and extract a small subset of information-rich features which can be used to distinguish between benign and malignant breast tumors. The data acquisition system consists of a commercially available ultrasonic imaging system (Octoson) interfaced to an RF amplifier with TGC and a Nicolet digital oscilloscope. A single-sideband receiver, under construction, will be used with a pulsed-Doppler transmitter and analog tape unit to acquire measures of spatial and temporal blood flow patterns associated with the tumor. Data from both systems are transmitted to a PDP 11/34 for offline processing. Feature extraction algorithms generate measurements related to image texture and Doppler related features. Over 150 features can be generated for each digitized A-scan. This high dimensional measurement space is then reduced to a subspace containing features whose information content iS sufficient to allow an accurate two-way classification of breast tumors

208

ABSTRACTS,

ULTFASONfC

IMAGING

AND

TISSUE

C~~~ER~Z~TION

SY~OSIUM

into benign and malignant categories. Truth values for these categories are determined by biopsy. Statistical and information theoretic measures used in dimensionality reduction are considered, as well as three different classification schemes. Preliminary results and the relative advantages/disadvantages of our approach will be discussed. OF ABDOMINAL ORGANS IN THIRTY NORMAL HUMAN SUBJECTS --IN VIVO DIFFERENTIATION BY THE ANALYSIS OF A-MODE ULTRASOUND WAVEFORMS, Joie Pierce Jones, Victor Gonzalez, Marty Behrens, and Leondard Ferrari, Department of Radiological Sciences, University of California, Irvine, CA 92717. This paper will describe the results of comprehensive measurements and analyses of rf A-line waveforms taken --in vivo on over 30 normal human subjects. All measurements were limited to abdominal organs and included liver, pancreas, and spleen. All measurements were accomplished by a computerized data analysis system designed for ultrasonic tissue characOne important feature of this system terization which we have implemented. is its ability to record both the phase and the amplitude of A-mode waveforms selected through regions of interest on a conventional B-mode ultrasonogram. Typically, we record 32 distinct lines which are 2048 points long (sampled at 20 MHz) through a specified region. This data acquisition process is accomplished within the time required to produce a conventional B-scan image. Various signal processing procedures have been applied to the data including amplitude histogram, intensity histogram, power spectrum, autocorrelation, cross correlation, real cepstrum, envelope correlation spectrum (ECS), and histogram of the first peaks in the autocorrelation of the power spectrum. Only the latter two techniques were successful Our results in differentiating between liver, pancreas, and spleen. suggest that tissue characterization by the analysis of A-mode ultrasound waveforms is still feasible. However, pathology specific algorithms may be required in which the algorithm development is based on the acoustical properties of the tissue or pathology in question. DIGITAL, ULTRASONIC, CEPSTRAL, AND SPECTRAL ANALYSIS OF --IN VITRO AND -IN VIVO LIVER AND KIDNEY, Frederic L. Lizzil, Donald L. King2, Ernest J. Feleppa', Nicholas Jaremkol, and Peter Wais, lRiverside Research Institute, New York, NY 10023, and 2Columbia University College of Physicians and Surgeons, New York, NY 10032. Digital, spectral, and cepstral analyses have been applied to fresh post-mortem samples of normal liver and kidney and abnormal kidney and to in vivo liver and kidney of normal and diseased volunteers. These analyses arebeing used to characterize normal and abnormal tissue, and therefore, to serve as a basis for diagnosing and evaluating disease and for selected and monitoring therapy. Ultrasonic rf echoes obtained directly from the transducers of an ATL ultrasonic scanner are digitized and transferred to a DEC PDP-11/55 computer for processing. Average power spectra are computed and normalized with reference to the spectrum of a glass-plate calibration target. Cepstra are computed from normalized power spectra (prior to averaging); averaging is applied to the cepstra. Spectral parameters of interest are the best-least-squares-fit slope (in dB/MHz~, average amplitude (in dB relative to the calibration spectrum) or intercept (of the s?ope line with the O-MHz axis), and scalloping (null spacing in MHz). Amplitude and intercept correlate with reflectivity. The distance between nulls in spectral scallops measures characteristic tissue spacings. Often, null separations in spectral scallops cannot be measured accurately, e.g., several scallop-patterns may coexist in a spectrum. In such cases, cepstra separate the individual scallop patterns and directly

209