IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control ,
vol. 61, no. 9,
September
2014
1489
Adaptive Motion Estimation of Shear Shock Waves in Soft Solids and Tissue With Ultrasound Gianmarco Pinton, Jean-Luc Gennisson, Mickaël Tanter, and François Coulouvrat Abstract—Shear shock waves in soft solids, such as in tissue, have different regions of complex motion that can change rapidly across a single wave profile, especially at the shock front. Conventional tracking algorithms are not well adapted to the task of simultaneously tracking the discontinuous shock front and smooth regions away from the shock. An adaptive algorithm based on the normalized cross-correlation and a correlation-weighted median filter is presented. The proposed adaptive algorithm combines two features: first, it adapts the window size to optimize the correlation value based on the deformation, and second, it rejects inaccurate estimates with a median-weighted filter. For simulated ultrasound data, where the displacements are known, it is shown that the estimated velocity error for the adaptive algorithm is less than 1/3 of the error for non-adaptive normalized cross-correlation. The addition of the weighted median filter to the adaptive algorithm significantly improves the shock tracking performance. The shock position and rise-time error is almost an order of magnitude better with the median-weighted filter. This algorithm is then used to track shock wave propagation with data acquired by a high-frame-rate ultrasound scanner in a tissue-mimicking agar and gelatin phantom. The shock front is not resolved with conventional algorithms but it is clearly visible with the proposed adaptive median-weighted algorithm.
I. Introduction
S
hear shock waves in soft solids were first observed in 2003 [1]. These observations were made possible by two developments: The first development was high-framerate ultrasound scanners that could acquire B-mode images at up to 10 000 frames/second [2]. Much like high-speed cameras, these ultrafast ultrasound scanners are capable of capturing physical behavior that is not observable at conventional frame-rates. The second development was techniques that can accurately determine motion from the RF ultrasound data [3]. High-frame-rate ultrasound in combination with RF motion tracking algorithms currently remains the only known method that can quantita-
Manuscript received February 10, 2014; accepted June 2, 2014. G. Pinton, J.-L. Gennisson, and F. Coulouvrat are with the Centre National de la Recherche Scientifique, Paris, France (e-mail: gfp@unc. edu). G. Pinton and F. Coulouvrat are with Institut Jean le Rond d’Alembert, Université Pierre et Marie Curie, UMR 7190, Paris, France. G. Pinton is currently with the Department of Biomedical Engineering, University of North Carolina at Chapel Hill and North Carolina State University, Chapel Hill, NC. J.-L. Gennisson and M. Tanter are with the Institut Langevin, École Supérieure de Physique et Chimie Industrielles, ParisTech, CNRS UMR 7587, INSERM U979, Paris, France. DOI http://dx.doi.org/10.1109/TUFFC.2014.3063 0885–3010
tively measure the propagation of shear waves at depth in optically opaque soft solids such as human tissue. The study of shock wave propagation in soft solids is a new research domain that could have a significant impact in trauma research such as in traumatic brain injuries or in thoracoabdominal trauma. We hypothesize that upon impact, shear waves propagate in soft tissue. Because of the large Mach numbers, the shear displacements can quickly generate shock fronts that may cause further injury. Displacement or velocity estimation is a key step in the conversion of RF ultrasound data to images of shear wave propagation and is the topic of this paper. Research in this domain has been driven principally by elasticity imaging, in which the mechanical properties of soft tissue are determined by ultrasonic measurements of tissue motion [4]. Tissue motion can be divided into two categories: transient or continuous waves, in which the motion is linear and the shear displacements are relatively small and smooth [5], [6], and quasistatic deformations in strain imaging, in which the displacements are large and discontinuous but no waves propagate [7], [8]. The motion estimation algorithms for transient wave elastography are tuned to measure small, sub-sample displacements with high correlation values. Tissue motion is generated by small amounts of ultrasound radiation force. This induces a maximum displacement that is typically in the range of 10 μm and the displacement between two frames is much smaller than this value [9]. Some of the first algorithms that were established to calculate these displacements are based on a phase shift estimate [10], [11]. They have the advantage of being fast and easy to implement, especially on in-phase and quadrature demodulated data. Another type of tracking method is based on finding the best match between a portion of the reference signal when compared with the delayed signal. The popularity of these pattern matching methods is due to the variety of functions that can be used to match the signals and the trade-offs they give in terms of computational efficiency and performance [9], [12]. One of the most commonly used functions is discrete normalized cross-correlation. To achieve a subsample resolution with a discrete function, these algorithms are coupled with various stages of interpolation. They tend to be less computationally efficient than phase shift calculations but generally offer better accuracy. In strain imaging, the tissue deformation is generated by pressing the ultrasound probe into the patient’s soft tissue. The displacements are typically much larger than
© 2014 IEEE
1490
IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control ,
in transient elastography—normally on the order of a few millimeters. These larger displacements result in lower correlation values. Furthermore, where there is a tissue interface, such as between folds of the brain, blood vessels, or fluid filled cysts, there is a slip boundary condition [13]. Consequently, the tissue motion can appear to be discontinuous between different regions. Because of the challenging conditions, the tracking algorithms used in strain elastography have evolved into adaptive multi-level or multi-scale schemes to improve the estimation stability [14]–[25]. Instead of performing an exhaustive search, an initial motion estimate can be used as initial guess for the following iteration [18]. The motion estimate can furthermore be guided by a quality indicator [26]. In one example of a non-iterative adaptive static elastography algorithm, the strain is estimated with normalized cross-correlation with different search window lengths [16]. By assuming that the strain is uniform, it was shown that an optimal window length could be calculated as a function of depth. This yielded significant improvements in the signal-tonoise ratio of the final strain estimates. Unlike static elastography, shear shock waves propagating in an inhomogeneous medium generate very heterogeneous displacement and strain fields that cannot be approximated a priori. The iteratively adaptive algorithms presented here differ significantly from previous adaptive algorithms in two ways. First, a continuous spline-based version of normalized cross-correlation is used because it has previously been demonstrated that it is more accurate than discrete normalized correlation [27]. The performance of this algorithm is compared with discrete normalized correlation with pre- and post-correlation interpolation [9]. Second, it establishes and uses a medianweighted filter that combines a filtering and thresholding operation. The design of an effective nonsmoothing filter is essential to preserve the high-frequency characteristics of shock fronts. It is shown that the proposed filter increases the accuracy without smoothing the shock front by simultaneously optimizing the spatial domain of dependence and correlation value for each motion estimate. The algorithm optimizes the search window length without an a priori knowledge of the displacements in the medium. The correlation-weighted median filter is compared with an unweighted median filter and a correlation threshold separately, to demonstrate the improvement. The performance of the adaptive algorithms is evaluated with synthetic RF signals that have a known displacement generated by an analytic description of a shear shock wave. It is shown that because the adaptive algorithms have a 2-D spatial domain of dependence, they are more effective in reducing the error of the shock position and rise time when compared with algorithms that only have a 1-D spatial domain of dependence. Once the error has been characterized, the algorithms are used to measure the shear shock wave propagation experimentally with data acquired by a high-frame-rate ultrasound scanner in a tissue-mimicking agar and gelatin phantom.
vol. 61, no. 9,
September
2014
Low-frame-rate (
vol. 61, no. 9,
September
2014
1489
Adaptive Motion Estimation of Shear Shock Waves in Soft Solids and Tissue With Ultrasound Gianmarco Pinton, Jean-Luc Gennisson, Mickaël Tanter, and François Coulouvrat Abstract—Shear shock waves in soft solids, such as in tissue, have different regions of complex motion that can change rapidly across a single wave profile, especially at the shock front. Conventional tracking algorithms are not well adapted to the task of simultaneously tracking the discontinuous shock front and smooth regions away from the shock. An adaptive algorithm based on the normalized cross-correlation and a correlation-weighted median filter is presented. The proposed adaptive algorithm combines two features: first, it adapts the window size to optimize the correlation value based on the deformation, and second, it rejects inaccurate estimates with a median-weighted filter. For simulated ultrasound data, where the displacements are known, it is shown that the estimated velocity error for the adaptive algorithm is less than 1/3 of the error for non-adaptive normalized cross-correlation. The addition of the weighted median filter to the adaptive algorithm significantly improves the shock tracking performance. The shock position and rise-time error is almost an order of magnitude better with the median-weighted filter. This algorithm is then used to track shock wave propagation with data acquired by a high-frame-rate ultrasound scanner in a tissue-mimicking agar and gelatin phantom. The shock front is not resolved with conventional algorithms but it is clearly visible with the proposed adaptive median-weighted algorithm.
I. Introduction
S
hear shock waves in soft solids were first observed in 2003 [1]. These observations were made possible by two developments: The first development was high-framerate ultrasound scanners that could acquire B-mode images at up to 10 000 frames/second [2]. Much like high-speed cameras, these ultrafast ultrasound scanners are capable of capturing physical behavior that is not observable at conventional frame-rates. The second development was techniques that can accurately determine motion from the RF ultrasound data [3]. High-frame-rate ultrasound in combination with RF motion tracking algorithms currently remains the only known method that can quantita-
Manuscript received February 10, 2014; accepted June 2, 2014. G. Pinton, J.-L. Gennisson, and F. Coulouvrat are with the Centre National de la Recherche Scientifique, Paris, France (e-mail: gfp@unc. edu). G. Pinton and F. Coulouvrat are with Institut Jean le Rond d’Alembert, Université Pierre et Marie Curie, UMR 7190, Paris, France. G. Pinton is currently with the Department of Biomedical Engineering, University of North Carolina at Chapel Hill and North Carolina State University, Chapel Hill, NC. J.-L. Gennisson and M. Tanter are with the Institut Langevin, École Supérieure de Physique et Chimie Industrielles, ParisTech, CNRS UMR 7587, INSERM U979, Paris, France. DOI http://dx.doi.org/10.1109/TUFFC.2014.3063 0885–3010
tively measure the propagation of shear waves at depth in optically opaque soft solids such as human tissue. The study of shock wave propagation in soft solids is a new research domain that could have a significant impact in trauma research such as in traumatic brain injuries or in thoracoabdominal trauma. We hypothesize that upon impact, shear waves propagate in soft tissue. Because of the large Mach numbers, the shear displacements can quickly generate shock fronts that may cause further injury. Displacement or velocity estimation is a key step in the conversion of RF ultrasound data to images of shear wave propagation and is the topic of this paper. Research in this domain has been driven principally by elasticity imaging, in which the mechanical properties of soft tissue are determined by ultrasonic measurements of tissue motion [4]. Tissue motion can be divided into two categories: transient or continuous waves, in which the motion is linear and the shear displacements are relatively small and smooth [5], [6], and quasistatic deformations in strain imaging, in which the displacements are large and discontinuous but no waves propagate [7], [8]. The motion estimation algorithms for transient wave elastography are tuned to measure small, sub-sample displacements with high correlation values. Tissue motion is generated by small amounts of ultrasound radiation force. This induces a maximum displacement that is typically in the range of 10 μm and the displacement between two frames is much smaller than this value [9]. Some of the first algorithms that were established to calculate these displacements are based on a phase shift estimate [10], [11]. They have the advantage of being fast and easy to implement, especially on in-phase and quadrature demodulated data. Another type of tracking method is based on finding the best match between a portion of the reference signal when compared with the delayed signal. The popularity of these pattern matching methods is due to the variety of functions that can be used to match the signals and the trade-offs they give in terms of computational efficiency and performance [9], [12]. One of the most commonly used functions is discrete normalized cross-correlation. To achieve a subsample resolution with a discrete function, these algorithms are coupled with various stages of interpolation. They tend to be less computationally efficient than phase shift calculations but generally offer better accuracy. In strain imaging, the tissue deformation is generated by pressing the ultrasound probe into the patient’s soft tissue. The displacements are typically much larger than
© 2014 IEEE
1490
IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control ,
in transient elastography—normally on the order of a few millimeters. These larger displacements result in lower correlation values. Furthermore, where there is a tissue interface, such as between folds of the brain, blood vessels, or fluid filled cysts, there is a slip boundary condition [13]. Consequently, the tissue motion can appear to be discontinuous between different regions. Because of the challenging conditions, the tracking algorithms used in strain elastography have evolved into adaptive multi-level or multi-scale schemes to improve the estimation stability [14]–[25]. Instead of performing an exhaustive search, an initial motion estimate can be used as initial guess for the following iteration [18]. The motion estimate can furthermore be guided by a quality indicator [26]. In one example of a non-iterative adaptive static elastography algorithm, the strain is estimated with normalized cross-correlation with different search window lengths [16]. By assuming that the strain is uniform, it was shown that an optimal window length could be calculated as a function of depth. This yielded significant improvements in the signal-tonoise ratio of the final strain estimates. Unlike static elastography, shear shock waves propagating in an inhomogeneous medium generate very heterogeneous displacement and strain fields that cannot be approximated a priori. The iteratively adaptive algorithms presented here differ significantly from previous adaptive algorithms in two ways. First, a continuous spline-based version of normalized cross-correlation is used because it has previously been demonstrated that it is more accurate than discrete normalized correlation [27]. The performance of this algorithm is compared with discrete normalized correlation with pre- and post-correlation interpolation [9]. Second, it establishes and uses a medianweighted filter that combines a filtering and thresholding operation. The design of an effective nonsmoothing filter is essential to preserve the high-frequency characteristics of shock fronts. It is shown that the proposed filter increases the accuracy without smoothing the shock front by simultaneously optimizing the spatial domain of dependence and correlation value for each motion estimate. The algorithm optimizes the search window length without an a priori knowledge of the displacements in the medium. The correlation-weighted median filter is compared with an unweighted median filter and a correlation threshold separately, to demonstrate the improvement. The performance of the adaptive algorithms is evaluated with synthetic RF signals that have a known displacement generated by an analytic description of a shear shock wave. It is shown that because the adaptive algorithms have a 2-D spatial domain of dependence, they are more effective in reducing the error of the shock position and rise time when compared with algorithms that only have a 1-D spatial domain of dependence. Once the error has been characterized, the algorithms are used to measure the shear shock wave propagation experimentally with data acquired by a high-frame-rate ultrasound scanner in a tissue-mimicking agar and gelatin phantom.
vol. 61, no. 9,
September
2014
Low-frame-rate (
