Locally optimized correlation-guided Bayesian adaptive regularization for ultrasound strain imaging

A cardiac strain imaging framework with adaptive Bayesian regularization (ABR) is proposed for in vivo assessment of murine cardiac function. The framework uses ultrasound (US) radio-frequency data collected with a high frequency (fc = 30MHz) imaging system and a multi-level block matching algorithm with ABR to derive inter-frame cardiac displacements. Lagrangian cardiac strain (radial, er and longitudinal, el) tensors were derived by segmenting the myocardial wall starting at the ECG R-wave and accumulating interframe deformations over a cardiac cycle. In vivo feasibility was investigated through a longitudinal study with two mice (one ischemia-perfusion (IR) injury and one sham) imaged at five sessions (pre-surgery (BL) and 1,2,7 and 14 days post-surgery). End-systole (ES) strain images and segmental strain curves were derived for quantitative evaluation. Both mice showed periodic variation of er and el strain at BL with segmental synchroneity. Infarcted regions of IR mouse at Day 14 were associated with reduced or sign reversed ES er and el values while the sham mouse had similar or higher strain than at BL. Infarcted regions identified in vivo were associated with increased collagen content confirmed with Masson's Trichrome stained ex vivo heart sections.Clinical Relevance-Higher quality cardiac strain images derived with RF data and Bayesian regularization can potentially improve the sensitivity and accuracy of non-invasive assessment of cardiovascular disease models.

Ultrasound strain imaging maps the tissue stiffness in the region of interest by estimating strain when the tissue is stressed. Prestress and poststress ultrasound echoes are processed to estimate the displacements, and the strain image is computed by a spatial derivative of the estimated displacement. Due to the nature of derivative operations, smaller errors in displacement estimation can cause large noise in the computed strain image. When tissue displacements are estimated using the phase of the autocorrelation between the prestress and poststress data (i.e., 1D autocorrelation method), it has been previously reported that local variations in ultrasonic center frequency due to speckle can introduce errors in the estimated displacement. We have developed a new method to compute strain based on two-dimensional autocorrelation. By estimating the local ultrasound frequency, we can improve the accuracy of displacement estimates and hence reduce the noise in strain images. We have analyzed the effect of local frequency changes on noise in strain images and the improvement in the strain signal-to-noise ratio with the 2D autocorrelation method. The simulation results are supported by experiments with homogenous gelatin phantoms and show that strain signal-to-noise ratio with 2D autocorrelation is consistently higher than that with 1D autocorrelation. The 2D autocorrelation can increase the strain signal-to-noise ratio by up to 200%, which leads us to believe that our estimation method can significantly improve the quality of strain images.

The rupture of atherosclerotic plaques represents one of the leading causes of morbidity and mortality. The aim of our study is to assess the morphology, functionality and composition of plaques formed in common carotid artery using a non-invasive in-vivo imaging technology, namely combined ultrasound, strain and photoacoustics. The imaging approach was tested ex-vivo on excised rabbit aortas (two with and two without plaques) using the integrated vascular imaging probe. The artery was flushed with blood at a pulsatile flow (1.5 Hz pulse rate and 50 m/s flow velocity). Ultrasound B-mode images and strain images, recorded at 296 frames per second, were obtained using a 12 MHz linear array transducer. For photoacoustic imaging the same ultrasound transducer was coupled with 19 optical fibers connected to OPO laser system (operated at 5 ns pulse duration, 590 nm and 700 nm spectral range). To analyze the advantages of the three imaging techniques, hyper- and hypoechogenic regions were separately isolated and the corresponding strain and photoacoustic signals were compared. Spatial variations in ultrasound echogeneicity are related to changes in arterial structure and function, i.e. hyperechogenic regions correspond to 15% or more decrease in peak strain rate and correlate with fibrotic tissue (as identified by photoacoustic images). The results of our study suggest that combined ultrasound, strain and photoacoustic imaging provides complementary information. Analysis of co-registered ultrasound, strain and photoacoustic images can be used to differentiate morphology, functionality and composition of plaques necessary for clinical diagnosis.

Since 2009, there have been ≥170 articles published in the cardiopulmonary literature on right ventricular (RV) echocardiographic strain imaging, including several in Circulation: Cardiovascular Imaging . Of particular interest has been the effort to use RV echocardiographic strain imaging to assess RV systolic function in pulmonary hypertension and the relationship of RV strain to disease severity and outcomes. In this issue of Circulation: Cardiovascular Imaging , Fine et al,1 from the Mayo Clinic echo group, report on retrospectively evaluated data on the relationship of RV strain to outcomes in the largest such study to date, encompassing 575 patients with known or suspected pulmonary hypertension without elevated pulmonary venous pressure during a median follow-up period of 16.5 months.1 The data show a strong association between reductions in mean longitudinal strain and all-cause mortality, and cardiopulmonary mortality and cardiopulmonary events. An exhaustive biostatistical analysis shows the relationship to be quantitative, stronger than that of other echo indices including tricuspid annular plane systolic excursion, independent of clinical covariates and resulting in a significant improvement in predictive value when combined with relevant clinical variables, as reflected in a significant increase in the integrated discrimination improvement index and the c-statistic for the primary end point. Thus, the results promise to improve clinical risk assessment significantly for such patients and may improve clinical decision making. Article see p 711 The apparent benefit is, in large part, a reflection of the relatively weak performance of older 2-dimensional (2D) echocardiographic indices of RV performance including fractional area change and tricuspid annular plane systolic excursion in this setting. RV imaging has been a minefield for echocardiography in adults since its inception. The complexity of ventricular shape, the thinness of the RV free wall, and the limitations in adults of transthoracic and subcostal windows in permitting full visualization …

We demonstrate the feasibility of ultrasound saline infused sonohysterography based strain imaging, for the determination of variations in the stiffness of uterine masses in-vivo. Strain images are estimated using a two-dimensional multilevel hybrid algorithm developed for ultrasound sector array transducers. Coarse displacements are initially estimated using envelope data, followed by guided finer displacement estimation using window lengths on the order of 4 wavelengths and 5A-lines on RF data. Strain images are obtained by estimating displacement slopes using least squares estimation. The strain images provide additional new information that may supplement current clinical B-mode imaging, used for diagnosis of uterine abnormalities. Stiffer masses such as fibroids are clearly differentiated on the strain images, which appear to be darker on the strain images with the characteristic decorrelation "halo" when compared to the normal uterine tissue. In addition, uterine polyps are visualized with a lighter grayscale when compared to the background myometrium, indicating the presence of a softer mass. Our results demonstrate the feasibility of improving clinical diagnosis based on strain imaging.

Performance of Two Dimensional Displacement and Strain Estimation Techniques Using a Phased Array Transducer

This study developed a robust algorithm for realtime ultrasonic strain imaging that employs dynamic programming techniques. Since tissue motion under external mechanical stimuli should be highly continuous, the speckle tracking problem fits well within the general framework of the hidden Markov model and can be solved as an optimization problem. A cost function combining correlation and motion continuity was used to regularize motion tracking. We found that the new algorithm provides more accurate displacement estimates than our previous algorithm for in vivo clinical data. In particular, the new algorithm is capable of tracking larger frame-average tissue deformation (1-2%) and increasing strain image consistency in a sequence of images. The new algorithm can also tolerate larger local strain (approx. 10%). Preliminary results also suggest that a significantly longer sequence of high contrast strain images (e.g. 45 vs. 15 in one cancer dataset) could be obtained with the new algorithm compared to the previous algorithm. We also achieved more than 10 frames/second with a 3 cm by 3 cm region of interest, which is sufficient to provide real-time feedback during in vivo elasticity imaging

Stroke and myocardial infarction are initiated by rupturing vulnerable atherosclerotic plaques. With noninvasive ultrasound elastography, these plaques might be detected in carotid arteries. However, since the ultrasound beam is generally not aligned with the radial direction in which the artery pulsates, radial and circumferential strains need to be derived from axial and lateral data. Conventional techniques to perform this conversion have the disadvantage that lateral strain is required. Since the lateral strain has relatively poor accuracy, the quality of the radial and circumferential strains is reduced. In this study, the radial and circumferential strain estimates are improved by combining axial strain data acquired at multiple insonification angles. Adaptive techniques to correct for grating lobe interference and other artifacts that occur when performing beam steering at large angles are introduced. Acquisitions at multiple angles are performed with a beam steered linear array. For each beam steered angle, there are two spatially restricted regions of the circular vessel cross section where the axial strain is closely aligned with the radial strain and two spatially restricted regions (different from the radial strain regions) where the axial strain is closely aligned with the circumferential strain. These segments with high quality strain estimates are compounded to form radial or circumferential strain images. Compound radial and circumferential strain images were constructed for a homogeneous vessel phantom with a concentric lumen subjected to different intraluminal pressures. Comparison of the elastographic signal-to-noise ratio (SNR(e)) and contrast-to-noise ratio (CNR(e)) revealed that compounding increases the image quality considerably compared to images from 0 degrees information only. SNR(e) and CNR(e) increase up to 2.7 and 6.6 dB, respectively. The highest image quality was achieved by projecting axial data, completed with a small segment determined by either principal component analysis or by application of a rotation matrix.

The common process of window-based ultrasound elastography (USE) is based on estimating the displacement fields and calculating the strain images by differentiating them. The displacement fields can be obtained using operators such as normalized cross-correlation (NCC). However, the provided displacements suffer from the artifacts caused by two dominant sources, the decorrelation, and amplitude modulation (AM) error. The main contribution of this study is to obtain strain images of higher quality and reduce the effect of artifacts. To this end, a weighted displacement estimation method is proposed. The proposed method consists of two steps. Some changes are applied to the windowing process, speeding up displacement estimation using the NCC method and filtering the provided displacements in two stages using median and weighted windows. We have assessed the performance of different weighted windows in conjunction with NCC on simulation and phantom data. This evaluation has not been implemented on the displacement fields yet. Thus, we tried to represent their improving effects on the estimated strains compared to the simple NCC approach and suggest the best window. The quantitative evaluation is performed in terms of elastographic signal to noise ratio (SNRe) and elastographic contrast to noise ratio (CNRe) on simulation and phantom data. The root mean square error (RMSE) comparison is also implemented for the simulation data. The visual and quantitative comparisons reveal that the proposed method substantially reduces artifacts, improves the strain image quality, and outperforms the NCC method.

A novel two-way image quality assessment method is proposed for free-hand strain imaging. In elasticity imaging, tissue with different stiffness exhibit varying contrast in the strain images and detectability of a lesion is measured using elastographic contrast-to-noise ratio (CNR <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">e</sub> ). Representing quality of strain images quantitatively is vital for improving imaging techniques and also for clinical diagnosis. It avoids the subjective approach of interpreting strain images. Conventionally, contrast between stiff lesion and surrounding soft tissue is measured using contrast-to-noise ratio and strain image with the highest CNR <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">e</sub> amplitude is considered an optimal strain image. However experimental results have suggested that merely CNR <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">e</sub> metric is often misleading and does not always represent the true elastic modulus contrast as the correlation coefficient falls below an acceptable levels and accuracy is compromised. Therefore in this study, the objective is to propose a comprehensive strain image quality assessment method which is reliable for clinical examinations and research.

Chapter 8 - Lateral and shear strain imaging for ultrasound elastography

In this study, a real-time compression based ultrasound elasticity imaging prototype was developed and evaluated for liver radio-frequency ablation (RFA) assessment under ex vivo and in vivo conditions. The prototype was implemented on a commercial ultrasound system (iU22, Philips Healthcare) for an L12-5 linear transducer and a C5-1 curvilinear transducer. Under the ex vivo condition, two bovine liver samples were used for L12-5 and C5-1 respectively. Under the in vivo condition, a pig was anesthetized and mechanically ventilated. Ultrasound elasticity imaging experiments were performed during breath-hold in order to reduce respiratory motion artifacts. Post-RFA strain images were compared against several imaging modalities including MRI, contrast-enhanced CT and gross pathology. For the ex vivo case, a Rita Starburst probe (AngioDynamics, Queensbury, NY) was used to perform RFA and induce tissue deformation for strain imaging. Real-time strain images were generated pre- and post-RFA. After the ablation, the liver was imaged by a Panorama 1-T MRI system (Philips Healthcare), then it was dissected along the approximate ultrasound imaging plane for lesion dimension measurement using a caliper. The post-RFA strain images depicted an ablation zone with increased stiffness that was well correlated with the physiological changes observed in the T1-weighted MR image and gross pathology. For the second ex vivo liver sample, the strain image revealed an ablation zone of 4.90 cm ? 2.85 cm, and the MR image and gross pathology showed a zone of 4.56 cm ? 3.24 cm and of 4.91 cm ? 3.65 cm, respectively. For the in vivo study, three different commercial RFA needles were employed: the Covidien (Mansfield, MA) CoolTip single prong and triple-cluster probes, and the Rita Starburst probe. The animal underwent contrast-enhanced CT imaging immediately after RFA. It was then sacrificed and the ablated liver lobes were dissected for ex vivo MRI. All three imaging modalities showed comparable measurement of lesion dimensions. The triple-cluster Covidien probe and the Rita probe with expanding tines seemed to perform better with elasticity assessment than the Covidien single prong probe. Preliminary results from this study demonstrate that real-time ultrasound strain imaging is potentially a valuable tool for assessing liver RFA. Further technical development of this prototype will advance toward real-time clinical use.

Strain imaging has been performed using 1D, 2D and 3D ultrasound data. Usually, strain is only estimated in the beam direction, because of the availability of phase information. However, strain estimates in other directions are also desired, since most tissues are anisotropic. Beam steering enables measurement of different projections of the 2D displacement field while using phase information. The full 2D displacement field and strain tensor can be derived using multiple projections. This study aims at deriving the full 2D strain tensor by estimating only displacement along the beam while steering the beam at three angles, i.e. 0?, a positive, and a negative beam steering angle. Displacements were iteratively estimated using 2D cross-correlations. The 0? acquisition yielded the vertical displacement vector. Projection of the other measured displacements yielded the horizontal (lateral) vector. The full 2D strain tensor was derived using 1D least squares strain estimators. Since horizontal displacements and strains are usually the least accurate, these were first focused on. Horizontal displacements were estimated for simulated radiofrequency (rf) data of a block with a four times stiffer cylindrical inclusion before and after application of 2% vertical compression. Root mean squared error (RMSE) of the horizontal displacements were calculated for every symmetric pair of beam steering angles in the range of -30 to 30? with angular increments of 5?. The RMSE analysis was performed for signal-to-noise (SNR) levels of 5, 10, 20 and 50 dB. Beam steering combinations of about -25? and 25? were found to be most optimal. The technique's performance was maintained regardless of the SNR level. For an SNR of 5 dB a reduction in RMSE of up to 70% was observed. In analogy with the simulations, experiments were performed with a similar inclusion phantom made of polyvinyl alcohol. Using data from beam steering angles of -25?, 0? and 25? strain images were constructed for all strain components and compared to strain images constructed without beam steering. The horizontal and shear strain images improved clearly with beam steering. Finally, the technique was also applied to derive radial and circumferential strain images for a homogeneous vessel phantom made of gelatin. Radially, SNR <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">e</sub> and CNR <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">e</sub> increased with 2.2 dB and 8 dB and circumferentially with 2.6 dB and 9.2 dB compared to non steered strain estimates. To conclude, the full 2D strain tensor can be accurately measured by using three beam steering angles.

Second-order ultrasound elastography (SOUL) is a regularized speckle-tracking algorithm with demonstrated promise to produce high-quality strain images. SOUL optimizes a cost function that consists of a data fidelity term and a regularization term. However, SOUL manually selects its regularization weights to generate strain images, which limits its efficiency and usability by making it highly user-dependent. We developed a deep learning approach to autonomously select the regularization weights required to reduce user dependency. A convolutional neural network (CNN) was trained on simulated datasets and fine-tuned using phantom data to classify a strain image to be acceptable or unacceptable and predict optimal parameters based on the strain image quality. When evaluated on simulated and public phantom datasets, the CNN achieved 100.00% recall with F1 scores ≥95%. Strain images from manual and automated tuning showed similar smoothness and contrast, highlighting the potential to eliminate manual SOUL parameter selection. The CNN-based autonomous selector reduces user dependence, enabling optimal strain imaging for inexperienced users and clinicians.

A 256 element toroidal piezocomposite transducer with a diameter of 70 mm and a radius of curvature of 70 mm working at 3 MHz was used. A 7.5 MHz ultrasound imaging probe was integrated to the centre of the HIFU transducer. This paper discusses the interest in using real‐time hand held elastography to guide a HIFU treatment. 16 lesions were produced and imaged intra‐operatively in vivo in 4 pigs. Strain images were obtained and displayed at a frame rate identical to that of the sonograms (23 frames/s) using the ultrasound imaging probe placed in the centre of the HIFU device. The compression was applied manually. At the end of the procedure the liver was removed and imaged using a 1.5 T MR‐scanner. The dimensions of the coagulated volumes measured on gross pathology were 22.1 mm (average diameter) and 19.8 mm (average depth). The contrast between ablated and untreated liver was higher on strain images (23 dB) than on the B‐Mode (3 dB), T1‐weighted (3 dB) and T2‐weighted images (13 dB). The correlation between dimensions measured on gross pathology and dimensions measured on B‐mode images, strain images, T1‐weighted and T2‐weighted images were r = 0.51, r = 0.68, r = 0.72 and r = 0.74 respectively. The correlation between dimensions measured on strain images and dimensions measured on T1‐weighted and T2‐weighted images were r = 0.77 and r = 0.78 respectively. In conclusion, the device presented in this study allows the creation of large ablations volumes over short periods of time. In addition, these in vivo results show that realtime hand‐held strain imaging performed with an ultrasound probe integrated to the HIFU device can be used intra‐operatively to visualize the dimensions of HIFU ablations with a precision similar to MRI, but with improved contrast.