Classification of Thermally Ablated Tissue Using Diagnostic Ultrasound

Abstract

In this work the application of an ultrasound based classification system for the detection of thermally ablated tissue is
presented. In-vitro and in-vivo experiments have been carried out using an RF ablation device. A conventional diagnostic ultrasound
device was used for measurements. Several features from spectral and spatial domain were extracted from ultrasound RF data
and processed by the classifier. Classification results are presented in binary coagulation maps

Full text

Page 1

CLASSIFICATION OF THERMALLY ABLATED
TISSUE USING DIAGNOSTIC ULTRASOUND

Stefan Siebers1, Ulrich Scheipers1, Johannes Hänsler2, Markus Frieser2,
Deike Strobel2, Christoph Welp3, Jürgen Werner3, Eckhart Hahn2and
Helmut Ermert1
1Institute of High Frequency Engineering, Ruhr-University, 44780 Bochum, Germany;
2Department of Internal Medicine, University Erlangen-Nürnberg, Ulmenweg 18, 91054
Erlangen, Germany; 3Institute of Biomedical Engineering, Ruhr-University, 44780 Bochum,
Germany
Abstract: In this work the application of an ultrasound based classification system for
the detection of thermally ablated tissue is presented. In-vitro and in-vivo
experiments have been carried out using an RF ablation device. A
conventional diagnostic ultrasound device was used for measurements. Several
features from spectral and spatial domain were extracted from ultrasound RF
data and processed by the classifier. Classification results are presented in
binary coagulation maps.
Key words: tissue characterization; classification; monitoring; RF ablation; thermal
therapy; coagulation.
1.
INTRODUCTION
Tissue characterization has been widely used throughout the ultrasound
research community for various applications1,2, amongst others for
differentiating pathological from healthy tissue. In this work, a classification
system for automated tissue characterization is investigated towards its
ability for monitoring thermal therapies as RF ablation. In this context, a
differentiation of thermally coagulated and noncoagulated or viable tissue is
desired. In several research projects, tissue characterizing ultrasound
parameters have been used for this purpose, however, in general a single

Page 2

2

parameter has been used on its own3,4,5. In this work, more than 40
parameters from spectral and spatial domain have been evaluated in a first
step, regarding their potential in detecting coagulated tissue. The best
combination has been processed by a linear classifier. The classification
result is presented in binary coagulation maps indicating the thermally
ablated zone.
Stefan Siebers, Ulrich Scheipers and Helmut Ermert
2.
EXPERIMENTS AND DATA ACQUISITION
2.1
RF Ablation
For RF Ablation a conventional device (Elektrotom HiTT 106, Integra
ME GmbH, Tuttlingen, Germany) was used. The system consists of a 60 W,
375 kHz RF generator to which needle and neutral electrode are connected,
and of a piston pump which delivers saline solution to the needle electrode
for open perfusion. Tissue impedance is measured by the device throughout
the application. Saline flow can be controlled automatically. If necessary, a
bolus of saline is given to compensate for variation in tissue impedance. The
hollow needle electrodes consist of isolated steel except for the 15mm active
tip. The active tip had 6 micro holes for infusion of saline solution into the
tissue. During ablation, temperature of up to 100°C are reached in the
surrounding of the needle tip, inducing thermal coagulations of up to 5 cm in
diameter.
2.2
Data Acquisition
Data were acquired using two conventional diagnostic ultrasound
systems, the Siemens Antares for in vitro measurements, and the Siemens
Elegra for in vivo measurements.
The Siemens Antares was equipped with the Siemens Axius Direct
Ultrasound Research Interface (URI), which provides RF data at 40 MHz
sampling frequency and 16 bit resolution. Using the Siemens Elegra,
baseband data were obtained accessing its internal operating system via
telnet and downloading the data via FTP. All features needed for controlling
the Elegra were provided by custom made software. With both systems, a
3.5 MHz curved array transducer was used.

Page 3

CLASSIFICATION OF THERMALLY ABLATED TISSUE

2.3
In Vitro Experiments
3
Samples of bovine liver have been used for in vitro experiments. The
samples were placed in a custom made holder. The needle electrode was
placed perpendicular to the imaging plane to avoid artifacts in the images.
The imaging depth was set to 60 mm. The power of the RF generator was set
to 20 W for 5 to 8 minutes, leading to relatively smooth coagulations without
larger gas bubbles that also might cause artifacts in the images. Ultrasound
RF data were acquired after the coagulation process and stored on a PC.
Afterwards the liver specimens were sliced in the imaging plane.
Photographs were taken to serve as a reference for later analysis.
2.4
In Vivo Experiments
All animal experiments were performed according to the German animal
protection law and official permission to perform the animal experiments
was obtained. 10 domestic pigs were included in this study. The anesthetized
animals were placed in supine position on the operating table. RF ablation
was performed intraoperatively. The livers were excised, sliced and also
photographed.
3.
DATA PROCESSING
3.1
Extraction of Tissue Characterizing Parameters
To obtain spatially resolved tissue characterizing parameter images, each
set of RF data was subdivided in up to 1700 overlapping ROIs. The size of
the ROIs was set to 128 samples (2.5 mm) axially and 16 lines (4 degrees)
laterally. Overlap was set to 50 %, both axially and laterally. For each ROI a
set of 40 spectral and textural features was calculated. Spectral features take
into account changes in spectral behavior of coagulated tissue, like increase
in attenuation.
To consider changes in spectral properties due to the coagulation process,
21 parameters from the frequency domain were calculated. Spectral
estimates were obtained using Fourier transform (FT) as well as
autoregressive modeling (AR) using the Burg method. The optimal order of
the AR-model was determined using the Akaike information criterion to
be 14.
As backscatter measures midband, slope and intercept values of a
straight-line fit to the spectrum were used. Frequency dependent attenuation

Page 4

4

coefficients were also used. The center frequency for attenuation estimation
was determined using spectral moments as well as a least square fit to the
spectrum. Estimates of attenuation coefficients were obtained using multi-
narrowband7 and spectral shift method8. Furthermore, estimated coefficients
of the autoregressive model of order 14 were used as tissue characterizing
parameters. All texture parameters were calculated after demodulating the
RF signals using Hilbert transform. Moreover, for the calculation of second
order texture parameters, image data were quantized to 64 intensity levels.
First order texture parameters account for variations in intensity without
spatial reference, second order texture parameters account for changes in the
spatial arrangement of intensities. First order parameters used in this study
were maximum, minimum, mean value, signal-to-noise ratio, squared signal-
to-noise ratio, standard deviation, contrast, entropy, kurtosis, skewness, and
full-width-at-half-maximum of intensities. Second order texture parameters
were extracted from normalized cooccurrence matrices6 and include contrast,
inverse difference moment, entropy, second angular moment, and
correlation.
Stefan Siebers, Ulrich Scheipers and Helmut Ermert
3.2
Classification
In a first step, features have been evaluated regarding their potential in
discriminating coagulated from noncoagulated tissue. For that purpose, the
coagulated and a noncoagulated region were manually delineated in the B-
Mode images according to the photographs and transferred to the parameter
images. Two groups of ROIs were created thereby, each representing
coagulated and noncoagulated tissue, respectively. The distribution of each
feature for both groups was obtained calculating histograms. Defining a
threshold kc, sensitivity and specificity and hence the receiver operating
characteristic (ROC) curve were determined for each feature. The area under
the ROC curve served as a quality criterion for selectivity of each feature.
The Maximum-Likelihood measure was used to classify each ROI
separately. Classification was done by leave-one-out validation over cases,
i.e. to classify one ROI the ROIs of the same case were not included in the
training data set. The classification results were presented in binary
coagulation maps. To evaluate the classification result, sensitivity and
specificity for each map were determined.
To find the best combination of features to be processed by a classifier,
all possible combinations of features must be tried, as a matter of principle.
Since this exhaustive scheme would call for an enormous computational
cost, an approximation for the best combination was alternatively found by
sequential forward selection. Initially, the feature with the largest area under
the ROC curve was processed separately by the classifier. Subsequently, the

Page 5

CLASSIFICATION OF THERMALLY ABLATED TISSUE

feature with the second largest area under the ROC curve was added. If the
sum of sensitivity and specificity improved with the new feature, it was
included for classification, otherwise it was excluded. This procedure was
repeated for all remaining features.
5
4.
RESULTS AND DISCUSSION
1 cm
1 cm
1 cm
a)b)
c)
coagulation

Figure 1. In vivo results. Photograph of coagulation (a), manually delineated coagulated
region (b), coagulation map obtained by classification system (c).
15 in vitro experiments were included in this study. The training data set
consisted of 1802 ROIs representing group 1 (coagulated), and 2092 ROIs
representing group 2 (noncoagulated). Sensitivity and specificity reached
0.9±0.12 and 0.84±0.09, respectively. The lower specificity is mainly due to
an overestimation of the coagulated zone, possibly caused by the size of the
ROIs. Future work will therefore incorporate the influence of the ROI size to
the classification result.
The training data set for in vivo experiments included 2122 ROIs
representing group 1 and 1388 ROIs representing group 2. Overall 10
datasets were acquired and classified. Sensitivity and specificity achieved
thereby were 0.86±0.11 and 0.76±0.09.

Page 6

6

Stefan Siebers, Ulrich Scheipers and Helmut Ermert
Classification results were displayed in scan converted binary images,
serving as coagulation maps. The maps can be superimposed to the B-Mode
image for better visualization of the coagulated region.
5.
ACKNOWLEDGEMENTS
This work is an activity of the Ruhr Center of Excellence for Medical
Engineering (Kompetenzzentrum Medizintechnik Ruhr, KMR) Bochum. It is
supported by the German Federal Ministry of Education and Research
(Bundesministerium für Bildung und Forschung), grant No. 13N8079, and
Siemens Medical Solutions, Siemens AG.
6.
REFERENCES
1. Scheipers U, Ermert H, Sommerfeld HJ, Garcia-Schürmann M, Senge T, Philippou S:
Ultrasonic multifeature tissue characterization for prostate diagnostics. Ultrasound Med
Biol 29(8): 1137-1149, 2003.
2. Siebers S, Geier B, Scheipers U, Vogt M, Mumme A, Ermert H: Classification of Venous
Thrombosis combining Ultrasound Elastography and Tissue Characterization. Proc 2004
IEEE Ultrason Symp, in press, 2004.
3. Bevan PD, Sherar MD: B-scan ultrasound imaging of thermal coagulation in bovine liver:
Frequency Shift attenuation mapping. Ultrasound Med Biol 27(6): 809-817, 2001.
4. Lemor R, Kleffner B, Tretbar S, Schmitt R: Ultrasound temperature and attenuation
monitoring for controlling the laser induced thermo therapy. Acoust. Imag., 25: 395-400,
2000
5. Righetti R, Kallel F, Stafford R, Price R, Krouskop T, Hazle J, Ophir J: Elastographic
Chracterization Of HIFU-Induced Lesions In Canine Livers. Ultrasound Med Biol 25(7):
1099-1113, 1999.
6. Haralick RM, Shanmugam K, Dinstein I: Textural Features For Image Classification. IEEE
Trans Syst Man Cybern, SMC-3(6), 1973
7. Cloostermans MJTM, Thijssen JM: A beam corrected estimation of the frequency
dependent attenuation of biological tissues from backscattered ultrasound. Ultrason
Imaging 5: 136-147, 1983.
8. Fink M, Hottier F, Cardoso JF: Ultrasonic signal processing for in vivo attenuation
measurement – short time fourier analysis. Ultrason Imaging 5: 117-135, 1983