Training Deep Network Ultrasound Beamformers With Unlabeled In Vivo Data.

Conventional delay-and-sum (DAS) beamforming is highly efficient but also suffers from various sources of image degradation. Several adaptive beamformers have been proposed to address this problem, including more recently proposed deep learning methods. With deep learning, adaptive beamforming is typically framed as a regression problem, where clean ground-truth physical information is used for training. Because it is difficult to know ground truth information in vivo, training data are usually simulated. However, deep networks trained on simulations can produce suboptimal in vivo image quality because of a domain shift between simulated and in vivo data. In this work, we propose a novel domain adaptation (DA) scheme to correct for domain shift by incorporating unlabeled in vivo data during training. Unlike classification tasks for which both input domains map to the same target domain, a challenge in our regression-based beamforming scenario is that domain shift exists in both the input and target data. To solve this problem, we leverage cycle-consistent generative adversarial networks to map between simulated and in vivo data in both the input and ground truth target domains. Additionally, to account for separate as well as shared features between simulations and in vivo data, we use augmented feature mapping to train domain-specific beamformers. Using various types of training data, we explore the limitations and underlying functionality of the proposed DA approach. Additionally, we compare our proposed approach to several other adaptive beamformers. Using the DA DNN beamformer, consistent in vivo image quality improvements are achieved compared to established techniques.

Evaluating Input Domain and Model Selection for Deep Network Ultrasound Beamforming.

Improving ultrasound B-mode image quality remains an important area of research. Recently, there has been increased interest in using deep neural networks (DNNs) to perform beamforming to improve image quality more efficiently. Several approaches have been proposed that use different representations of channel data for network processing, including a frequency-domain approach that we previously developed. We previously assumed that the frequency domain would be more robust to varying pulse shapes. However, frequency- and time-domain implementations have not been directly compared. In addition, because our approach operates on aperture domain data as an intermediate beamforming step, a discrepancy often exists between network performance and image quality on fully reconstructed images, making model selection challenging. Here, we perform a systematic comparison of frequency- and time-domain implementations. In addition, we propose a contrast-to-noise ratio (CNR)-based regularization to address previous challenges with model selection. Training channel data were generated from simulated anechoic cysts. Test channel data were generated from simulated anechoic cysts with and without varied pulse shapes, in addition to physical phantom and in vivo data. We demonstrate that simplified time-domain implementations are more robust than we previously assumed, especially when using phase preserving data representations. Specifically, 0.39- and 0.36-dB median improvements in in vivo CNR compared to DAS were achieved with frequency- and time-domain implementations, respectively. We also demonstrate that CNR regularization improves the correlation between training validation loss and simulated CNR by 0.83 and between simulated and in vivo CNR by 0.35 compared to DNNs trained without CNR regularization.

On the Effects of Constitutive Properties and Roughness of a Hard Inclusion in Soft Tissue on B-mode Images.

We perform finite element modeling of pulse-echo ultrasound of a hard inclusion in a soft tissue to gain a better understanding of B-mode image brightness characteristics. We simulate a pressure wave emitted by an ultrasound transducer through the inclusion-tissue medium by prescribing suitable boundary conditions, and collect the scattered wave response to simulate the behavior of the transducer array used for pulse-echo ultrasound. We form B-mode images from simulated channel data using standard delay and sum beamforming. We establish the accuracy of the finite element model by comparing the point spread function with that obtained from Field II ultrasound simulation program. We also demonstrate qualitative validation by comparing the brightness characteristics of rough and smooth surfaced circular inclusions with experimental images of a cylindrical metal tool immersed in a water tank. We next conduct simulation studies to evaluate changes in B-mode image brightness intensity and contrast related to different constitutive properties, namely, compressibility of the inclusion, impedance contrast between the host and inclusion, and surface roughness of the inclusion. We find that the intensity observed behind a hard inclusion in the axial direction is strongly affected by the compressibility and roughness of the inclusion. Also, the perceived width of the stone based on the intensity is greater for rougher stones. Our study indicates that imaging of compressible inclusions may benefit from targeted B-mode image forming algorithms. Our modeling framework can potentially be useful in differentiating hard inclusions from surrounding parenchyma, and for classifying kidney stones or gallstones.

Independent Component-Based Spatiotemporal Clutter Filtering for Slow Flow Ultrasound.

Effective tissue clutter filtering is critical for non-contrast ultrasound imaging of slow blood flow in small vessels. Independent component analysis (ICA) has been considered by other groups for ultrasound clutter filtering in the past and was shown to be superior to principal component analysis (PCA)-based methods. However, it has not been considered specifically for slow flow applications or revisited since the onset of other slow flow-focused advancements in beamforming and tissue filtering, namely angled plane wave beamforming and full spatiotemporal singular value decomposition (SVD) (i.e., PCA-based) tissue filtering. In this work, we aim to develop a full spatiotemporal ICA-based tissue filtering technique facilitated by plane wave applications and compare it to SVD filtering. We compare ICA and SVD filtering in terms of optimal image quality in simulations and phantoms as well as in terms of optimal correlation to ground truth blood signal in simulations. Additionally, we propose an adaptive blood independent component sorting and selection method. We show that optimal and adaptive ICA can consistently separate blood from tissue better than principal component analysis (PCA)-based methods using simulations and phantoms. Additionally we demonstrate initial in vivo feasibility in ultrasound data of a liver tumor.

A Spatial Coherence Beamformer Design for Power Doppler Imaging.

Acoustic clutter is a primary source of image degradation in ultrasound imaging. In the context of flow imaging, tissue and acoustic clutter signals are often much larger in magnitude than the blood signal, which limits the sensitivity of conventional power Doppler in SNR-limited environments. This has motivated the development of coherence-based beamformers, including Coherent Flow Power Doppler (CFPD), which have demonstrated efficacy in mitigating sources of diffuse clutter. However, CFPD uses a measure of normalized coherence, which incurs a non-linear relationship between image intensity and the magnitude of the blood echo. As a result, CFPD is not a robust approach to study gradation of blood signal energy, which depicts the fractional moving blood volume. We propose the application of mutual intensity, rather than normalized coherence, to retain the clutter suppression capability inherent in coherence beamforming, while preserving the underlying signal energy. Feasibility of this approach was shown via Field II simulations, phantoms, and in vivo human liver data. In addition, we derive an adaptive statistical threshold for the suppression of residual noise signals. Overall, this beamformer design shows promise as an alternative technique to depict flow volume gradation in cluttered imaging environments.

Non-contrast power Doppler ultrasound imaging for early assessment of trans-arterial chemoembolization of liver tumors.

Trans-arterial chemoembolization (TACE) is an important yet variably effective treatment for management of hepatic malignancies. Lack of response can be in part due to inability to assess treatment adequacy in real-time. Gold-standard contrast enhanced computed tomography and magnetic resonance imaging, although effective, suffer from treatment-induced artifacts that prevent early treatment evaluation. Non-contrast ultrasound is a potential solution but has historically been ineffective at detecting treatment response. Here, we propose non-contrast ultrasound with recent perfusion-focused advancements as a tool for immediate evaluation of TACE. We demonstrate initial feasibility in an 11-subject pilot study. Treatment-induced changes in tumor perfusion are detected best when combining adaptive demodulation (AD) and singular value decomposition (SVD) techniques. Using a 0.5 s (300-sample) ensemble size, AD + SVD resulted in a 7.42 dB median decrease in tumor power after TACE compared to only a 0.06 dB median decrease with conventional methods.

Combining Slow Flow Techniques With Adaptive Demodulation for Improved Perfusion Ultrasound Imaging Without Contrast.

Noncontrast perfusion ultrasound imaging remains challenging due to spectral broadening of the tissue clutter signal caused by patient and sonographer hand motion. To address this problem, we previously introduced an adaptive demodulation scheme to suppress the bandwidth of tissue prior to high-pass filtering. Our initial implementation used single plane wave power Doppler imaging and a conventional tissue filter. Recent advancements in beamforming and tissue filtering have been proposed for improved slow flow imaging, including coherent flow power Doppler (CFPD) imaging and singular value decomposition (SVD) filtering. Here, we aim to evaluate adaptive demodulation in conjunction with improvements in beamforming and filtering using simulations, single-vessel phantoms, and an in vivo liver tumor embolization study. We show that simulated blood-to-background contrast-to-noise ratios are highest when using adaptive demodulation with CFPD and a 100-ms ensemble, which resulted in a 13.6-dB average increase in contrast-to-noise ratio compared to basic IIR filtering alone. We also show that combining adaptive demodulation with SVD and with CFPD + SVD results in 9.3- and 19-dB increases in contrast-to-noise ratios compared to IIR filtering alone at 700- and 500-ms ensembles for phantom data with 1- and 5-mm/s average flows, respectively. In general, combining techniques resulted in higher signal-to-noise, contrast-to-noise, and generalized contrast-to-noise ratios in both simulations and phantoms. Finally, adaptive demodulation with SVD resulted in the largest qualitative and quantitative changes in tumor-to-background contrast postembolization.

In vitro feasibility of next generation non-linear beamforming ultrasound methods to characterize and size kidney stones.

Ultrasound imaging for kidney stones suffers from poorer sensitivity, diminished specificity, and overestimation of stone size compared to computed tomography (CT). The purpose of this study was to demonstrate in vitro feasibility of novel ultrasound imaging methods comparing traditional B-mode to advanced beamforming techniques including plane wave synthetic focusing (PWSF), short-lag spatial coherence (SLSC) imaging, mid-lag spatial coherence (MLSC) imaging with incoherent compounding, and aperture domain model image reconstruction (ADMIRE). The ultrasound techniques were evaluated using a research-based ultrasound system applied to an in vitro kidney stone model at 4 and 8 cm depths. Stone diameter sizing and stone contrast were compared among the different techniques. Analysis of variance was used to analyze the differences among group means, with p < 0.05 considered significant, and a Student's t test was used to compare each method with B-mode, with p < 0.0025 considered significant. All stones were detectable with each method. MLSC performed best with stone sizing and stone contrast compared to B-mode. On average, B-mode sizing error ± SD was > 1 mm (1.2 ± 1.1 mm), while those for PWSF, ADMIRE, and MLSC were < 1 mm (- 0.3 ± 2.9 mm, 0.6 ± 0.8, 0.8 ± 0.8, respectively). Subjectively, MLSC appeared to suppress the entire background thus highlighting only the stone. The ADMIRE and SLSC techniques appeared to highlight the stone shadow relative to the background. The detection and sizing of stones in vitro are feasible with advanced beamforming methods with ultrasound. Future work will include imaging stones at greater depths and evaluating the performance of these methods in human stone formers.

Feasibility of non-linear beamforming ultrasound methods to characterize and size kidney stones.

PURPOSE: Ultrasound methods for kidney stone imaging suffer from poor sensitivity and size overestimation. The study objective was to demonstrate feasibility of non-linear ultrasound beamforming methods for stone imaging, including plane wave synthetic focusing (PWSF), short-lag spatial coherence (SLSC) imaging, mid-lag spatial coherence (MLSC) imaging with incoherent compounding, and aperture domain model image reconstruction (ADMIRE). MATERIALS AND METHODS: The ultrasound techniques were evaluated in an in vitro kidney stone model and in a pilot study of 5 human stone formers (n = 6 stones). Stone contrast, contrast-to-noise ratio (CNR), sizing, posterior shadow contrast, and shadow width sizing were compared among the different techniques and to B-mode. CT imaging within 60 days was considered the gold standard stone size. Paired t-tests using Bonferroni correction were performed to evaluate comparing each technique with B-mode. RESULTS: Mean CT measured stone size was 6.0mm (range 2.9-12.2mm) with mean skin-to-stone distance 10.2cm (range 5.4-16.3cm). Compared to B-mode, stone contrast was best with ADMIRE (mean +12.2dB), while SLSC and MLSC showed statistically improved CNR. Sizing was best with ADMIRE (mean +1.3mm error), however this was not significantly improved over B-mode (+2.4mm). PWSF performed similarly to B-mode for stone contrast, CNR, SNR, and stone sizing. In the in vitro model, the shadow contrast was highest with ADMIRE (mean 10.5 dB vs 3.1 dB with B-mode). Shadow sizing was best with SLSC (mean error +0.9mm ± 2.9), however the difference compared to B-mode was not significant. CONCLUSIONS: The detection and sizing of stones are feasible with advanced beamforming methods with ultrasound. ADMIRE, SLSC, and MLSC hold promise for improving stone detection, shadow contrast, and sizing.

Model-based beamforming with plane wave synthesis in medical ultrasound.

We are interested in examining how our model-based beamforming algorithm, referred to as aperture-domain model image reconstruction (ADMIRE), performs on plane wave sequences in conjunction with synthetic aperture beamforming. We also aim to identify the impact of ADMIRE applied before and after synthetic focusing. We employed simulated phantoms using Field II and tissue-mimicking phantoms to evaluate ADMIRE as applied to synthetic sequencing. We generated plane wave images with and without synthetic aperture focusing (SAF) and measured contrast and contrast-to-noise ratio (CNR). For simulated cyst images formed from single plane waves, the contrast for delay-and-sum (DAS) and ADMIRE are 15.64 and 28.34 dB, respectively, whereas the CNR are 1.76 and 3.90 dB, respectively. We also applied ADMIRE to simulated resolution phantoms having a point target at 3 cm depth on-axis. We simulated the point spread functions from data obtained from 1 plane wave and 75 steered plane waves, along with linear scans with 3 and 4 cm- focal depths. We then compared the outcome of applying ADMIRE before and after SAF using 3 and 11 steered plane waves. Finally, we applied this to an in vivo carotid artery. Based on the findings in this study, ADMIRE can be adapted to full field insonification sequences to improve image quality in plane wave imaging. Additionally, we investigated how robustly ADMIRE performs in the presence of random noise. We then address identified limitations using a conventional envelope detection method with decluttered signals.

Adaptive Clutter Demodulation for Non-Contrast Ultrasound Perfusion Imaging.

Conventional Doppler ultrasound is useful for visualizing fast blood flow in large resolvable vessels. However, frame rate and tissue clutter caused by movement of the patient or sonographer make visualizing slow flow with ultrasound difficult. Patient and sonographer motion causes spectral broadening of the clutter signal, which limits ultrasound's sensitivity to velocities greater than 5-10 mm/s for typical clinical imaging frequencies. To address this, we propose a clutter filtering technique that may increase the sensitivity of Doppler measurements to at least as low as 0.52 mm/s. The proposed technique uses plane wave imaging and an adaptive frequency and amplitude demodulation scheme to decrease the bandwidth of tissue clutter. To test the performance of the adaptive demodulation method at suppressing tissue clutter bandwidths due to sonographer hand motion alone, six volunteer subjects acquired data from a stationary phantom. Additionally, to test in vivo feasibility, arterial occlusion and muscle contraction studies were performed to assess the efficiency of the proposed filter at preserving signals from blood velocities 2 mm/s or greater at a 7.8 MHz center frequency. The hand motion study resulted in initial average bandwidths of 175 Hz (8.60mm/s), which were decreased to 10.5 Hz (0.52 mm/s) at -60 dB using our approach. The in vivo power Doppler studies resulted in 4.73 dB and 4.80 dB dynamic ranges of the blood flow with the proposed filter and 0.15 dB and 0.16 dB dynamic ranges of the blood flow with a conventional 50 Hz high-pass filter for the occlusion and contraction studies, respectively.

A model and regularization scheme for ultrasonic beamforming clutter reduction.

Acoustic clutter produced by off-axis and multipath scattering is known to cause image degradation, and in some cases these sources may be the prime determinants of in vivo image quality. We have previously shown some success addressing these sources of image degradation by modeling the aperture domain signal from different sources of clutter, and then decomposing aperture domain data using the modeled sources. Our previous model had some shortcomings including model mismatch and failure to recover B-Mode speckle statistics. These shortcomings are addressed here by developing a better model and by using a general regularization approach appropriate for the model and data. We present results with L1 (lasso), L2 (ridge), and L1/L2 combined (elastic-net) regularization methods. We call our new method aperture domain model image reconstruction (ADMIRE). Our results demonstrate that ADMIRE with L1 regularization, or weighted toward L1 in the case of elastic-net regularization, have improved image quality. L1 by itself works well, but additional improvements are seen with elastic-net regularization over the pure L1 constraint. On in vivo example cases, L1 regularization showed mean contrast improvements of 4.6 and 6.8 dB on fundamental and harmonic images, respectively. Elastic net regularization (α = 0.9) showed mean contrast improvements of 17.8 dB on fundamental images and 11.8 dB on harmonic images. We also demonstrate that in uncluttered Field II simulations the decluttering algorithm produces the same contrast, contrast-tonoise ratio, and speckle SNR as normal B-mode imaging, demonstrating that ADMIRE preserves typical image features.