FNPC-SAM: Uncertainty-Guided False Negative/Positive Control for SAM on Noisy Medical Images.

The Segment Anything Model (SAM) is a recently developed all-range foundation model for image segmentation. It can use sparse manual prompts such as bounding boxes to generate pixel-level segmentation in natural images but struggles in medical images such as low-contrast, noisy ultrasound images. We propose a refined test-phase prompt augmentation technique designed to improve SAM's performance in medical image segmentation. The method couples multi-box prompt augmentation and an aleatoric uncertainty-based false-negative (FN) and false-positive (FP) correction (FNPC) strategy. We evaluate the method on two ultrasound datasets and show improvement in SAM's performance and robustness to inaccurate prompts, without the necessity for further training or tuning. Moreover, we present the Single-Slice-to-Volume (SS2V) method, enabling 3D pixel-level segmentation using only the bounding box annotation from a single 2D slice. Our results allow efficient use of SAM in even noisy, low-contrast medical images. The source code has been released at: https://github.com/MedICL-VU/FNPC-SAM.

Methods for Enhancing the Robustness of the Generalized Contrast-to-Noise Ratio.

The generalized contrast-to-noise ratio (gCNR) is a new but increasingly popular metric for measuring lesion detectability due to its use of probability distribution functions that increase robustness against transformations and dynamic range alterations. The value of these kinds of metrics has become increasingly important as it becomes clear that traditional metrics can be arbitrarily boosted with advanced beamforming or the right kinds of postprocessing. The gCNR works well for most cases; however, we will demonstrate that for some specific cases the implementation of gCNR using histograms requires careful consideration, as histograms can be poor estimates of probability density functions (PDFs) when designed improperly. This is demonstrated with simulated lesions by altering the amount of data and the number of bins used in the calculation, as well as by introducing some extreme transformations that are represented poorly by uniformly spaced histograms. In this work, the viability of a parametric gCNR implementation is tested, more robust methods for implementing histograms are considered, and a new method for estimating gCNR using empirical cumulative distribution functions (eCDFs) is shown. The most consistent methods found were to use histograms on rank-ordered data or histograms with variable bin widths, or to use eCDFs to estimate the gCNR.

A Coded Excitation Framework for High SNR Transcranial Ultrasound Imaging.

Robust transcranial ultrasound imaging is difficult due to poor image quality. In particular, low signal-to-noise ratio (SNR) limits sensitivity to blood flow and has hindered clinical translation of transcranial functional ultrasound neuroimaging thus far. In this work, we present a coded excitation framework to increase SNR in transcranial ultrasound without negatively impacting frame rate or image quality. We applied this coded excitation framework in phantom imaging and showed SNR gains as large as 24.78 dB and signal-to-clutter ratio gains as high as 10.66 dB with a 65 bit code. We also analyzed how imaging sequence parameters can impact image quality and showed how coded excitation sequences can be designed to maximize image quality for a given application. In particular, we show that considering the number of active transmit elements and the transmit voltage is critical for coded excitation with long codes. Finally, we applied our coded excitation technique in transcranial imaging of ten adult subjects and showed an average SNR gain of 17.91 ± 0.96 dB without a significant increase in clutter using a 65 bit code. We also performed transcranial power Doppler imaging in three adult subjects and showed contrast and contrast-to-noise ratio improvements of 27.32 ± 8.08 dB and 7.25 ± 1.61 dB, respectively with a 65 bit code. These results show that transcranial functional ultrasound neuroimaging may be possible using coded excitation.

Spatiotemporal Coherence to Quantify Sources of Image Degradation in Ultrasonic Imaging.

Thermal noise and acoustic clutter signals degrade ultrasonic image quality and contribute to unreliable clinical assessment. When both noise and clutter are prevalent, it is difficult to determine which one is a more significant contributor to image degradation because there is no way to separately measure their contributions in vivo. Efforts to improve image quality often rely on an understanding of the type of image degradation at play. To address this, we derived and validated a method to quantify the individual contributions of thermal noise and acoustic clutter to image degradation by leveraging spatial and temporal coherence characteristics. Using Field II simulations, we validated the assumptions of our method, explored strategies for robust implementation, and investigated its accuracy and dynamic range. We further proposed a novel robust approach for estimating spatial lag-one coherence. Using this robust approach, we determined that our method can estimate the signal-to-thermal noise ratio (SNR) and signal-to-clutter ratio (SCR) with high accuracy between SNR levels of -30 to 40 dB and SCR levels of -20 to 15 dB. We further explored imaging parameter requirements with our Field II simulations and determined that SNR and SCR can be estimated accurately with as few as two frames and sixteen channels. Finally, we demonstrate in vivo feasibility in brain imaging and liver imaging, showing that it is possible to overcome the constraints of in vivo motion using high-frame rate M-Mode imaging.

Multidimensional Clutter Filtering of Aperture Domain Data for Improved Blood Flow Sensitivity.

Singular value decomposition (SVD) is a valuable factorization technique used in clutter rejection filtering for power Doppler imaging. Conventionally, SVD is applied to a Casorati matrix of radio frequency data, which enables filtering based on spatial or temporal characteristics. In this article, we propose a clutter filtering method that uses a higher order SVD (HOSVD) applied to a tensor of aperture data, e.g., delayed channel data. We discuss temporal, spatial, and aperture domain features that can be leveraged in filtering and demonstrate that this multidimensional approach improves sensitivity toward blood flow. Further, we show that HOSVD remains more robust to short ensemble lengths than conventional SVD filtering. Validation of this technique is shown using Field II simulations and in vivo data.

Histogram Matching for Visual Ultrasound Image Comparison.

The widespread development of new ultrasound image formation techniques has created a need for a standardized methodology for comparing the resulting images. Traditional methods of evaluation use quantitative metrics to assess the imaging performance in specific tasks, such as point resolution or lesion detection. Quantitative evaluation is complicated by unconventional new methods and nonlinear transformations of the dynamic range of data and images. Transformation-independent image metrics have been proposed for quantifying task performance. However, clinical ultrasound still relies heavily on visualization and qualitative assessment by expert observers. We propose the use of histogram matching to better assess differences across image formation methods. We briefly demonstrate the technique using a set of sample beamforming methods and discuss the implications of such image processing. We present variations of histogram matching and provide code to encourage the application of this method within the imaging community.

Assessing the Robustness of Frequency-Domain Ultrasound Beamforming Using Deep Neural Networks.

We study training deep neural network (DNN) frequency-domain beamformers using simulated and phantom anechoic cysts and compare to training with simulated point target responses. Using simulation, physical phantom, and in vivo scans, we find that training DNN beamformers using anechoic cysts provided comparable or improved image quality compared with training DNN beamformers using simulated point targets. The proposed method could also be adapted to generate training data from in vivo scans. Finally, we evaluated the robustness of DNN beamforming to common sources of image degradation, including gross sound speed errors, phase aberration, and reverberation. We found that DNN beamformers maintained their ability to improve image quality even in the presence of the studied sources of image degradation. Overall, the results show the potential of using DNN beamforming to improve ultrasound image quality.

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.

Deep Neural Networks for Ultrasound Beamforming.

We investigate the use of deep neural networks (DNNs) for suppressing off-axis scattering in ultrasound channel data. Our implementation operates in the frequency domain via the short-time Fourier transform. The inputs to the DNN consisted of the separated real and imaginary components (i.e. in-phase and quadrature components) observed across the aperture of the array, at a single frequency and for a single depth. Different networks were trained for different frequencies. The output had the same structure as the input and the real and imaginary components were combined as complex data before an inverse short-time Fourier transform was used to reconstruct channel data. Using simulation, physical phantom experiment, and in vivo scans from a human liver, we compared this DNN approach to standard delay-and-sum (DAS) beamforming and an adaptive imaging technique that uses the coherence factor. For a simulated point target, the side lobes when using the DNN approach were about 60 dB below those of standard DAS. For a simulated anechoic cyst, the DNN approach improved contrast ratio (CR) and contrast-to-noise (CNR) ratio by 8.8 dB and 0.3 dB, respectively, compared with DAS. For an anechoic cyst in a physical phantom, the DNN approach improved CR and CNR by 17.1 dB and 0.7 dB, respectively. For two in vivo scans, the DNN approach improved CR and CNR by 13.8 dB and 9.7 dB, respectively. We also explored methods for examining how the networks in this paper function.

Improving Displacement Signal-to-Noise Ratio for Low-Signal Radiation Force Elasticity Imaging Using Bayesian Techniques.

Radiation force-based elasticity imaging is currently being investigated as a possible diagnostic modality for a number of clinical tasks, including liver fibrosis staging and the characterization of cardiovascular tissue. In this study, we evaluate the relationship between peak displacement magnitude and image quality and propose using a Bayesian estimator to overcome the challenge of obtaining viable data in low displacement signal environments. Displacement data quality were quantified for two common radiation force-based applications, acoustic radiation force impulse imaging, which measures the displacement within the region of excitation, and shear wave elasticity imaging, which measures displacements outside the region of excitation. Performance as a function of peak displacement magnitude for acoustic radiation force impulse imaging was assessed in simulations and lesion phantoms by quantifying signal-to-noise ratio (SNR) and contrast-to-noise ratio for varying peak displacement magnitudes. Overall performance for shear wave elasticity imaging was assessed in ex vivo chicken breast samples by measuring the displacement SNR as a function of distance from the excitation source. The results show that for any given displacement magnitude level, the Bayesian estimator can increase the SNR by approximately 9 dB over normalized cross-correlation and the contrast-to-noise ratio by a factor of two. We conclude from the results that a Bayesian estimator may be useful for increasing data quality in SNR-limited imaging environments.

Robust Tracking of Small Displacements With a Bayesian Estimator.

Radiation-force-based elasticity imaging describes a group of techniques that use acoustic radiation force (ARF) to displace tissue to obtain qualitative or quantitative measurements of tissue properties. Because ARF-induced displacements are on the order of micrometers, tracking these displacements in vivo can be challenging. Previously, it has been shown that Bayesian-based estimation can overcome some of the limitations of a traditional displacement estimator such as normalized cross-correlation (NCC). In this work, we describe a Bayesian framework that combines a generalized Gaussian-Markov random field (GGMRF) prior with an automated method for selecting the prior's width. We then evaluate its performance in the context of tracking the micrometer-order displacements encountered in an ARF-based method such as ARF impulse (ARFI) imaging. The results show that bias, variance, and mean-square error (MSE) performance vary with prior shape and width, and that an almost one order-of-magnitude reduction in MSE can be achieved by the estimator at the automatically selected prior width. Lesion simulations show that the proposed estimator has a higher contrast-to-noise ratio but lower contrast than NCC, median-filtered NCC, and the previous Bayesian estimator, with a non-Gaussian prior shape having better lesion-edge resolution than a Gaussian prior. In vivo results from a cardiac, radio-frequency ablation ARFI imaging dataset show quantitative improvements in lesion contrast-to-noise ratio over NCC as well as the previous Bayesian estimator.

Quantitative Imaging Assessment of an Alternative Approach to Surgical Mitral Valve Leaflet Resection: An Acute Porcine Study.

This study reports the initial in vivo use of a combined radiofrequency ablation and cryo-anchoring (RFC) catheter as an alternative to surgical mitral valve (MV) leaflet resection. Radiofrequency ablation thermally shrinks enlarged collagenous tissues, providing an alternative to leaflet resection, and cryo-anchoring provides reversible attachment of a catheter to freely mobile MV leaflets. Excised porcine MVs (n = 9) were tested in a left heart flow simulator to establish treatment efficacy criteria. Resected leaflet area was quantified by tracking markers on the leaflet surface, and leaflet length reductions were directly measured on echocardiography. Leaflet area decreased by 38 ± 2.7%, and leaflet length decreased by 9.2 ± 1.8% following RFC catheter treatment. The RFC catheter was then tested acutely in healthy pigs (n = 5) under epicardial echocardiographic guidance, open-chest without cardiopulmonary bypass, using mid-ventricular free wall access. Leaflet length was quantified using echocardiography. Quantitative assessment of MV leaflet length revealed that leaflet resection was successful in 4 of 5 pigs, with a leaflet length reduction of 13.3 ± 4.6%. Histological, mechanical, and gross pathological findings also confirmed that RFC catheter treatment was efficacious. The RFC catheter significantly reduces MV leaflet size in an acute animal model, providing a possible percutaneous alternative to surgical leaflet resection.

Short-lag spatial coherence imaging on matrix arrays, part II: Phantom and in vivo experiments.

In Part I of the paper, we demonstrated through simulation the potential of volumetric short-lag spatial coherence (SLSC) imaging to improve visualization of hypoechoic targets in three dimensions. Here, we demonstrate the application of volumetric SLSC imaging in phantom and in vivo experiments using a clinical 3-D ultrasound scanner and matrix array. Using a custom single-channel acquisition tool, we collected partially beamformed channel data from the fully sampled matrix array at high speeds and created matched Bmode and SLSC volumes of a vessel phantom and in vivo liver vasculature. 2-D and 3-D images rendered from the SLSC volumes display reduced clutter and improved visibility of the vessels when compared with their B-mode counterparts. We use concurrently acquired color Doppler volumes to confirm the presence of the vessels of interest and to define the regions inside the vessels used in contrast and contrast-to-noise ratio (CNR) calculations. SLSC volumes show higher CNR values than their matched B-mode volumes, while the contrast values appear to be similar between the two imaging methods.

Synthetic aperture focusing for short-lag spatial coherence imaging.

It has been demonstrated that short-lag spatial coherence (SLSC) ultrasound imaging can provide improved speckle SNR and lesion CNR compared with conventional Bmode images, especially in the presence of noise and clutter. Application of the van Cittert-Zernike theorem predicts that coherence among the ultrasound echoes received across an array is reduced significantly away from the transmit focal depth, leading to a limited axial depth of field in SLSC images. Transmit focus throughout the field of view can be achieved using synthetic aperture methods to combine multiple transmit events into a single final image. A synthetic aperture can be formed with either focused or diverging transmit beams. We explore the application of these methods to form synthetically focused channel data to create SLSC images with an extended axial depth of field. An analytical expression of SLSC image brightness through depth is derived for the dynamic receive focus case. Experimental results in a phantom and in vivo are presented and compared with dynamic receive focused SLSC images, demonstrating improved SNR and CNR away from the transmit focus and an axial depth of field four to five times longer.

Acoustic radiation force impulse imaging of mechanical stiffness propagation in myocardial tissue.

Acoustic radiation force impulse (ARFI) imaging has been shown to be capable of imaging local myocardial stiffness changes throughout the cardiac cycle. Expanding on these results, the authors present experiments using cardiac ARFI imaging to visualize and quantify the propagation of mechanical stiffness during ventricular systole. In vivo ARFI images of the left ventricular free wall of two exposed canine hearts were acquired. Images were formed while the heart was externally paced by one of two electrodes positioned on the epicardial surface and either side of the imaging plane. Two-line M-mode ARFI images were acquired at a sampling frequency of 120 Hz while the heart was paced from an external stimulating electrode. Two-dimensional ARFI images were also acquired, and an average propagation velocity across the lateral field of view was calculated. Directions and speeds of myocardial stiffness propagation were measured and compared with the propagations derived from the local electrocardiogram (ECG), strain, and tissue velocity measurements estimated during systole. In all ARFI images, the direction of myocardial stiffness propagation was seen to be away from the stimulating electrode and occurred with similar velocity magnitudes in either direction. When compared with the local epicardial ECG, the mechanical stiffness waves were observed to travel in the same direction as the propagating electrical wave and with similar propagation velocities. In a comparison between ARFI, strain, and tissue velocity imaging, the three methods also yielded similar propagation velocities.

A method for direct localized sound speed estimates using registered virtual detectors.

Accurate sound speed estimates are desirable in a number of fields. In an effort to increase the spatial resolution of sound speed estimates, a new method is proposed for direct measurement of sound speed between arbitrary spatial locations. The method uses the sound speed estimator developed by Anderson and Trahey. Their least squares fit of the received waveform's curvature provides an estimate of the wave's point of origin. The point of origin and the delay profile calculated from the fit are used to arrive at a spatially registered virtual detector. Between a pair of registered virtual detectors, a spherical wave is propagated. By beamforming the data, the time-of-flight between the two virtual sources can be calculated. From this information, the local sound speed can be estimated. Validation of the estimator is made using phantom and simulation data. The set of test phantoms consisted of both homogeneous and inhomogeneous media. Several different inhomogeneous phantom configurations were used for the physical validation. The simulation validation focused on the limits of target depth and signal-to-noise ratio on virtual detector registration. The simulations also compare the impact of two- and three-layer inhomogeneous media. The phantom results varied based on signal-to-noise ratio and geometry. The results for all cases were generally less than 1% mean error and standard deviation. The simulation results varied somewhat with depth and F/#, but primarily, they varied with signal-to-noise ratio and geometry. With two-layer geometries, the algorithm has a worst-case spatial registration bias of 0.02%. With three-layer geometries, the axial registration error gets worse with a bias magnitude up to 2.1% but is otherwise relatively stable over depth. The stability over depth of the bias in a given medium still allows for accurate sound speed estimates with a mean relative error less than 0.2%.

Effect of prior probability quality on biased time-delay estimation.

When properly constructed, biased estimators are known to produce lower mean-square errors than unbiased estimators. A biased estimator for the problem of ultrasound time-delay estimation was recently proposed. The proposed estimator incorporates knowledge of adjacent displacement estimates into the final estimate of a displacement. This is accomplished by using adjacent estimates to create a prior probability on the current estimate. Theory and simulations are used to investigate how the prior probability impacts the final estimate. The results show that with estimation quality on the order of the Cramer-Rao lower bound at adjacent locations, the local estimate in question should generally exceed the Cramer-Rao lower-bound limitations on performance of an unbiased estimator. The results as a whole provide additional confidence for the proposed estimator.

In vivo liver tracking with a high volume rate 4D ultrasound scanner and a 2D matrix array probe.

The effectiveness of intensity-modulated radiation therapy (IMRT) is compromised by involuntary motion (e.g. respiration, cardiac activity). The feasibility of processing ultrasound echo data to automatically estimate 3D liver motion for real-time IMRT guidance was previously demonstrated, but performance was limited by an acquisition speed of 2 volumes per second due to hardware restrictions of a mechanical linear array probe. Utilizing a 2D matrix array probe with parallel receive beamforming offered increased acquisition speeds and an opportunity to investigate the benefits of higher volume rates. In vivo livers of three volunteers were scanned with and without respiratory motion at volume rates of 24 and 48 Hz, respectively. Respiration was suspended via voluntary breath hold. Correlation-based, phase-sensitive 3D speckle tracking was applied to consecutively acquired volumes of echo data. Volumes were omitted at fixed intervals and 3D speckle tracking was re-applied to study the effect of lower scan rates. Results revealed periodic motion that corresponded with the heart rate or breathing cycle in the absence or presence of respiration, respectively. For cardiac-induced motion, volume rates for adequate tracking ranged from 8 to 12 Hz and was limited by frequency discrepancies between tracking estimates from higher and lower frequency scan rates. Thus, the scan rate of volume data acquired without respiration was limited by the need to sample the frequency induced by the beating heart. In respiratory-dominated motion, volume rate limits ranged from 4 to 12 Hz, interpretable from the root-mean-squared deviation (RMSD) from tracking estimates at 24 Hz. While higher volume rates yielded RMSD values less than 1 mm in most cases, lower volume rates yielded RMSD values of 2-6 mm.

Short-lag spatial coherence of backscattered echoes: imaging characteristics.

Conventional ultrasound images are formed by delay-and-sum beamforming of the backscattered echoes received by individual elements of the transducer aperture. Although the delay-and-sum beamformer is well suited for ultrasound image formation, it is corrupted by speckle noise and challenged by acoustic clutter and phase aberration. We propose an alternative method of imaging utilizing the short-lag spatial coherence (SLSC) of the backscattered echoes. Compared with matched B-mode images, SLSC images demonstrate superior SNR and contrast-to-noise ratio in simulated and experimental speckle-generating phantom targets, but are shown to be challenged by limited point target conspicuity. Matched B-mode and SLSC images of a human thyroid are presented. The challenges and opportunities of real-time implementation of SLSC imaging are discussed.