出席國際學術會議心得報告 出席國際學術會議心得報告 出席國際學術會議心得報告
計畫編號 NSC 97-2320-B-011-001-MY2
計畫名稱 基於方向性明亮度差異值之醫用超音波影像斑點雜訊抑制技術
出國人員姓名
服務機關及職稱 沈哲州 國立台灣科技大學電機系 副教授
會議時間地點 越南 胡志明市 2010/01/10~2010/01/14
會議名稱 The 3rd international conference on the development of biomedical engineering in Vietnam
發表論文題目
三、參加會議經過
本人利用國科會的補助經費應邀參加義 大利羅馬舉行的 The 3rd international
conference on the development of biomedical engineering in Vietnam 研討會，該研討會邀 請了數位在醫工領域的頂尖學者擔任講 者，其中尤以南加大熊克平教授(Shung, K.
Kirk)的高頻超音波影像與我的領域最為接 近，熊教授亦為本人 2003 接受教育部千里 馬計劃補助博生出國時的指導老師。在熊老 師本次的演講中知道了該研究團隊在高頻 聲學上的進展，尤其以 acoustic tweezer 的實驗成功最為令人振奮，也就是可以藉由 聲束能量對如脂質微粒產生拖拉效果，如果
未來可進一步順利運用在臨床上將可有很多實際的應用，如美容生技等即為一個例子!
Contrast Improvement by Combining Pulse Inversion with EMD and EEMD
Ai-Ho Liao1, Che-Chou Shen2 and Pai-Chi Li1,3
1Department of Electrical Engineering, National Taiwan University, Taipei, Taiwan, 2 Department of Electrical Engineering, National Taiwan University of Science and Technology, Taipei, Taiwan,
3Graduate Institute of Biomedical Electronics and Bioinformatics, National Taiwan University, Taipei, Taiwan
Abstract - Ultrasound nonlinear contrast imaging often has limited contrast because acoustic wave propagation in tissue is also nonlinear. In this study, a new ultrasound contrast imaging method which combines pulse-inversion technique with empirical mode decomposition (EMD) and the ensemble empirical mode decomposition (EEMD) in Hilbert-Huang transform (HHT) is explored to enhance the contrast between bubbles and surrounding tissues.
The EMD is a method associated with HHT that allows decomposition of the data into a finite number of intrinsic modes (IMF). EEMD method further utilizes the ensemble of IMF with added white noise to better differentiate the nonlinear microbubble signal from the nonlinear tissue signal. This hypothesis was tested on pulse-inversion (PI) nonlinear imaging. Our results showed that the contrast-to-tissue ratio (CTR) in the fundamental band and the second harmonic band was improved by 10.2 and 4.3 dB after EEMD. However, echoes from contrast agent areas after decomposition may vary from the original data. By extracting microbubble component and suppressing background noise, it is demonstrated both numerically and experimentally that the EEMD can noticeably improve image contrast.
I. INTRODUCTION
The use of a gas-encapsulated ultrasound contrast agent in ultrasound contrast imaging produces a strong backscattered signal due to the acoustic impedance mismatch between blood and the gas inside the microbubbles [1]. However, linear B-mode imaging is generally inadequate for distinguishing between tissue and blood because the enhancement of signals from blood may reduce the blood-to-tissue contrast. On the other hand, second-harmonic B-mode imaging is based on nonlinear properties of the contrast agent when the frequency of the impinging sound wave is near the resonance frequency of the constituent microbubbles [2].
Although the contrast is generally higher for second-harmonic imaging than for linear imaging, the contrast-to-tissue ratio (CTR) is still limited by the significant nonlinear response of the tissue associated with finite amplitude distortion [3], especially at a high mechanical index. Another alternative method is based on the subharmonic response of the microbubbles, because the tissue’s subharmonic response is minimal at acoustic pressures currently used in diagnostic ultrasound [3]. However, this requires the use of narrowband
transmit signals, which adversely affects the axial resolution. Moreover, harmonic power Doppler provides a higher signal-to-noise ratio [4], but the clutter signal generated by tissue motion may affect the detectability of low flows. All of these methods extract the harmonic signal by bandpass or high-pass filtering, and they suffer from spectral leakage resulting from a large transmit bandwidth [5].
The pulse-inversion (PI) technique is an alternative method to conventional filter-based harmonic and subharmonic imaging for extracting the nonlinear contrast signal [6]. The PI technique involves transmitting pairs of pulses with equal amplitudes but opposite polarities. The echoes elicited by the two pulses are then summed to cancel out the linear response of the medium. The imaging frequency determines whether PI imaging operates at the fundamental or second-harmonic frequency [7]. For example, a previous study using Levovist® demonstrated that the CTR can increase by up to 20 dB higher in fundamental PI imaging than in conventional linear imaging [6]. The present study built on that previous study by investigating the efficacy of using empirical mode decomposition (EMD) and ensemble empirical mode decomposition (EEMD) in the Hilbert-Huang transform (HHT) to further improve the detection of contrast agent in PI imaging.
The HHT is designed specifically for analyzing nonlinear and nonstationary signals [8]. It generally consists of two steps: EMD and Hilbert spectral analysis.
EMD is an adaptive time–frequency analysis method to resolve signals into several intrinsic mode functions (IMFs) that express the physical characteristics of the signals. One major drawback of EMD is the scale (frequency) mixing either consisting of a signal with widely disparate scales or a signal with a similar scale resides in different IMF components. This is alleviated by modifying the EMD method to the EEMD method, which considers an ensemble of data decompositions with added white noise and treats the resultant mean as the final true result [9]. Adding white noise has the effect of presenting a uniform reference frame in the time–frequency (time-scale) space and providing natural filter windows for the signals of a comparable scale to collate in a single component. This essentially eliminates the scale (frequency) mixing problem in EMD. EEMD fully utilizes statistical characteristics of the white noise to perturb the data in their true solution neighborhood, and then cancels out the white noise via ensemble averaging.
287
978-1-4244-4390-1/09/$25.00 ©2009 IEEE 2009 IEEE International Ultrasonics Symposium Proceedings 10.1109/ULTSYM.2009.0071
A. Empirical Mode Decomposition
EMD extracts the oscillatory mode that exhibits the highest local frequency values from the data, leaving the remainder as a residual. The starting point of EMD is identifying all local maxima and minima of the signals, which are separately connected by cubic spline curves to form the upper and lower envelopes, respectively. The mean of the two envelopes is calculated and subtracted from the original signal. Then, the subtracted signal is defined as the IMF function and the residue which is obtained and is used as the new input signal to repeat the sifting process. This process is stopped when the difference in the standard deviation of successive estimates of the IMF function or the residue becomes smaller than a predetermined threshold, or when the residue becomes a monotonic function from which no more IMFs can be extracted.
A signal y(t) decomposed with EMD can be expressed as the sum of cj and the last residue rn as In other words, we achieve a decomposition of the data into n empirical modes and a residue, rn, which can be either the mean trend or a constant. The decomposed components cj(t) contain the nonlinear oscillation components of the original ultrasound signal. Thus, it is possible to improve the CTR by decomposing the received nonlinear echoes into different IMFs and using the IMFs for imaging.
B. Ensemble Empirical Mode Decomposition
A noise-assisted data analysis method (EEMD) has been developed to overcome the scale mixing problem in EMD. The EEMD method consists of an ensemble of data decompositions with added white noise, and then treats the resultant mean as the final true result. The principle of EEMD is to add white noise, which populates the whole time-frequency space uniformly with the constituent components of different scales separated by a filter bank.
C. Simulations
The model used to simulate the tissue harmonic signal is obtained from numerical solutions of the Khokhlov–Zabolotskaya–Kuznetsov (KZK) nonlinear parabolic wave equation. The tissue signals are based on the time-domain solution to the KZK equation, which is an approximation of the nonlinear sound field produced by a piston source. The KZK equation can be written as
P P
a is the radius of the piston. The piston diameter was assumed to be 25.4 mm and the focal length was 70 mm.For the tissue signal, the transmit pulse has a pressure amplitude of 20 kPa at the transducer surface. The first, second, and third terms on the right-hand side of (4) correspond to the attenuation, nonlinear distortion, and diffraction, respectively. Pressure is normalized in the KZK equation as P = p/p0, where p is actual pressure and p0 is the pressure on the piston surface. The term r0 = ω0a2/2c0 is the Rayleigh distance, and ld = ρ0c03/βω0p0 is the shock-formation distance, where β = 1 + B/2A is a parameter describing the nonlinearity of the propagation medium. The nonlinear properties of water are approximated by making β = 3.5. The time variable τ
= ω0(t – z/c0) represents dimensionless retarded time.∇2⊥ is the Laplace operator, σ = z/r0 is the dimensionless axial coordinate and ξ = x/a is the dimensionless transverse coordinate.
For the contrast agent, a bubble-simulation tool (BubbleSim) was used to calculate the echo from a single microbubble (the tool was downloaded from the website of the IEEE UFFC society: www.ieee-uffc.org). The instantaneous radius of a bubble is approximated numerically by solving the Rayleigh-Plesset equation for an arbitrary impinging acoustic wave. Echoes from the bubble can be calculated from the bubble radius, wall velocity, and wall acceleration [7]. The simulations assumed that Sonovue (Bracco Diagnostics, Milan, Italy) was used, which has a shell thickness of 4 nm, a shell shear modulus of 46 MPa, and a shell viscosity of 0.25 Pa*s [7]. To simulate the nonlinear echo signal from the contrast agent, focal waveforms in the KZK simulations were used as the driving acoustic waveforms in the bubble simulator. Note also that two simulations were performed to acquire the sound fields corresponding to the positive and negative pulses in order to derive the summation signal in the PI technique.
D. Experiments
Fig. 1 shows a schematic diagram of the experimental setup for repetitive A-line measurements (i.e., M-mode) of the harmonic signal. PI observations were made using 60 repetitive A-line measurements. After EMD and EEMD, the CTRs were measured with the different IMFs.
The system consists of a 3.5-MHz single-element transducer that is used for both transmission and reception, and whose diameter and focal length are 25.4 mm and 70 mm, respectively. An arbitrary-function generator was used to generate the wide-band transmit Gaussian pulse at 2.5 MHz. The bandwidth of the transmit pulse was 80%. The waveform was then sent to a power amplifier to drive the transmit transducer. Note that the transducer has a −6 dB fractional bandwidth of about 80% so that both the corresponding fundamental and second-harmonic echoes can be received. A speckle-generating phantom containing a uniform distribution of glass beads was used as the tissue background in the image. A tubular void with a diameter of about 7 mm was fabricated inside the phantom as a microbubble container. A pulse repetition interval of 400 μs was used between the positive pulse and the negative II. IMAGING PRINCIPLES AND METHODS
288 2009 IEEE International Ultrasonics Symposium Proceedings
pulse, thus, the change of microbubbles in the sample volume between the two transmit pulses can be neglected between consecutive A-line measurements. To construct the M-mode image, the pulse pair was transmitted once for every 2.4 ms.
Fig. 1 Schematic diagram of the experimental setup for measuring A-line signals.
The Sonovue microbubbles used in the experiments were prepared according to the manufacturer’s instructions and diluted to a volume concentration of 1.5%. Sonovue microbubbles have a mean diameter of 3 µm, their resonance frequency is about 1.5–2 MHz. The gain of the power amplifier is kept low so as to prolong the life of the encapsulated bubbles. The acoustic pressure was measured with a PVDF needle hydrophone, and the MI at the focus was kept below 0.13. The received signals were then sent to an ultrasonic receiver.
Finally, an analog-to-digital converter was used to digitize the signal at 14-bit resolution and a sampling rate of 20 Msamples/s. The second-harmonic signal was filtered out by low-pass filtering the demodulated echo signal with a 1-MHz cutoff frequency.
III. RESULTS
The original simulated tissue and mixed tissue-microbubble PI signals and the sifting of IMF1–5 components with EMD are shown in Fig. 2 (left and right columns, respectively). After sifting, the tissue and the mixed tissue-microbubble PI signals exhibit very different characteristics starting from the second component (IMF2). In Fig. 3(a), note that, while there are some mixed tissue-microbubble residual signal remained at the odd-order harmonic frequencies (i.e. 2.5MHz and 7.5 MHz) after PI summation, the corresponding tissue signal is significantly cancelled. However, the tissue and mixed tissue-microbubble spectra closely resemble each other at the second-harmonic frequency (5 MHz). In other words, the original PI signals would exhibit better contrast between tissue and bubbles at the fundamental frequency than at the second-harmonic frequency.
Fig. 2. Original tissue harmonic and IMF1–5 components
of simulated PI signals (left column) and original mixed tissue-microbubble harmonic signal and IMF1–5 components of simulated PI signals (right column). All amplitudes are normalized to the maximum original amplitude for the corresponding tissue or mixed tissue-microbubble signal.
On the other hand, for the IMFs after sifting, the tissue and the mixed tissue-microbubble PI signals exhibit very different characteristics starting from the second component (IMF2). For example, the amplitude of the tissue signal appears to be suppressed in IMF2–5 relative to that of the mixed tissue-microbubble signal. The simulated PI spectra with IMF2 are shown in Fig. 3(b) for the tissue signal (solid line) and the mixed tissue-microbubble signal (dashed line). The figure clearly demonstrates that the mixed tissue-microbubble signal with IMF2 markedly differs from the tissue signal both at the fundamental and second-harmonic frequencies.
Fig. 3. (a) Spectra from the original tissue harmonic (solid line) and mixed tissue-microbubble (dashed line) simulated PI signals. (b) Spectra from the IMF2 component. All amplitudes are normalized to the fundamental amplitude for the corresponding tissue or mixed tissue-microbubble PI signal.
The intensities of the individual A-lines were then integrated over the slow-time axis (i.e., across transmit pulses). Fig. 4 shows the integrated intensities of PI A-line measurements for both fundamental and second-harmonic imaging. IMF1–3 were extracted by the EMD algorithm. The microbubble and tissue regions were at depths from 1 to 5 mm and from –3 to –7 mm, respectively. Here the CTR is estimated as the ratio of the averaged intensities between the bubble and tissue regions. In Fig. 4(a), the CTR in fundamental imaging is 18.7 dB for the original integrated PI A-line and 23.4 dB for IMF1. In Fig. 4(b), the CTR in second-harmonic imaging is 14.5 dB for the original integrated PI A-line and 14.9 dB for IMF1. With EMD, the CTR in fundamental and second-harmonic PI imaging increases by 4.7 (p-value<0.001) and 0.4 dB (p-value<0.001) for IMF1, respectively. On the other hand, the CTRs at other IMFs decrease both at the fundamental and second-harmonic frequencies. These results show that the CTR changes with the IMF, possibly improving at only certain IMF components, and that the CTR improvement using EMD appears to be greater at the fundamental frequency.
289 2009 IEEE International Ultrasonics Symposium Proceedings
Fig. 4 Integrated PI A-lines for IMF1 in fundamental imaging (a) and second-harmonic imaging (b) extracted by the EMD algorithm.
Fig. 5 shows the integrated PI A-line based on the same data with EEMD for fundamental and second-harmonic imaging. The CTR for IMF1 is 28.9 dB, in fundamental imaging (Fig. 5(a)), and the CTR for IMF3 is 18.8 dB in second-harmonic imaging (Fig. 5(b)).
With EEMD, the CTR increases by 10.2 dB (p-value<0.001) for IMF1 in fundamental imaging and by 4.3 dB (p-value<0.001) for IMF3 in secondharmonic imaging. The CTR for IMF3 in second-harmonic imaging is significantly higher for EEMD than for EMD.
Fig. 5 Integrated PI A-lines for IMF1and IMF3 in fundamental imaging (a) and second-harmonic imaging (b) extracted by the EEMD algorithm.
IV. DISCUSSION
Fig. 2 demonstrates that, for certain IMF components, EMD can effectively suppress the amplitude of the tissue signal while extracting the nonlinear oscillations of microbubbles. Consequently, utilizing certain IMFs when imaging potentially improves the CTR. Comparing the spectra in Fig. 3(a) and 3(b) also reveals that the signal decomposition further separates the nonlinear responses from tissue and microbubbles. The spectra in Fig. 3(a) show that the nonlinear response of the tissue closely resembles that of the bubbles especially at the second-harmonic frequency. On the other hand, Fig. 3(b) shows that EMD markedly improves the separation between tissue and bubbles at both the fundamental and second-harmonic frequencies.
Our experimentally integrated A-lines have also demonstrated that EMD leads to variation of CTR values among different IMF components and that the CTR improves only for certain IMF components. Note that since the PI second-harmonic signal is generated in both the tissue background and the bubble region, there is pronounced frequency mixing between the tissue and bubble signals. Consequently, the EMD method is less effective in improving the CTR of second-harmonic imaging due to its weakness in decomposing signals in the presence of frequency mixing.
Removing the problem of frequency mixing by utilizing the EEMD algorithm markedly improves the CTR in both PI fundamental imaging and PI second-harmonic imaging. Compared with EMD, the CTR in second-harmonic imaging is significantly higher for EEMD since EEMD is effective at separating the tissue harmonic signal and contrast harmonic signal of a similar scale. Therefore, EEMD is more suitable for imaging methods where there is significant overlap between the spectral components of the bubble and tissue signals.
V. CONCLUSIONS
The efficacies of the EMD and EEMD methods in improving the contrast in PI imaging was investigated. It was demonstrated both numerically and experimentally that EEMD can noticeably improve the image contrast in both PI fundamental and second-harmonic imaging.
However, the quality of the resultant image can be compromised due to the fragmentation of the speckle characteristics in the region containing contrast bubbles.
REFERENCES
[1] B. B. Goldberg, J. B. Liu, and F. Forsberg, “Ultrasound contrast agents: a review,” Ultrasound Med. Biol., vol. 20, no. 4, pp. 319–333,
[1] B. B. Goldberg, J. B. Liu, and F. Forsberg, “Ultrasound contrast agents: a review,” Ultrasound Med. Biol., vol. 20, no. 4, pp. 319–333,