Spatiotemporal Deep Learning-Based Cine Loop Quality Filter for Handheld Point-of-Care Echocardiography.

The reliability of automated image interpretation of point-of-care (POC) echocardiography scans depends on the quality of the acquired ultrasound data. This work reports on the development and validation of spatiotemporal deep learning models to assess the suitability of input ultrasound cine loops collected using a handheld echocardiography device for processing by an automated quantification algorithm (e.g. ejection fraction estimation). POC echocardiograms (n=885 DICOM cine loops from 175 patients) from two sites were collected using a handheld ultrasound device and annotated for image quality at the frame-level. Attributes of high-quality frames for left ventricular (LV) quantification included a temporally-stable LV, reasonable coverage of LV borders, and good contrast between the borders and chamber. Attributes of low-quality frames included temporal instability of the LV and/or imaging artifacts (e.g., lack of contrast, haze, reverberation, acoustic shadowing). Three different neural network architectures were investigated - (a) frame-level convolutional neural network (CNN) which operates on individual echo frames (VectorCNN), (b) single-stream sequence-level CNN which operates on a sequence of echo frames (VectorCNN+LSTM) and (c) two-stream sequence-level CNNs which operate on a sequence of echo and optical flow frames (VectorCNN+LSTM+Average, VectorCNN+LSTM+MinMax, and VectorCNN+LSTM+ConvPool). Evaluation on a sequestered test dataset containing 76 DICOM cine loops with 16,914 frames showed that VectorCNN+LSTM can effectively utilize both spatial and temporal information to regress the quality of an input frame (accuracy: 0.925, sensitivity = 0.860, specificity = 0.952), compared to the frame-level VectorCNN that only utilizes spatial information in that frame (accuracy: 0.903, sensitivity = 0.791, specificity = 0.949). Furthermore, an independent sample t-test indicated that the cine loops classified to be of adequate quality by the VectorCNN+LSTM model had a statistically significant lower bias in the automatically estimated EF (mean bias = - 3.73 ± 7.46 %, versus a clinically obtained reference EF) compared to the loops classified as inadequate (mean bias = -15.92 ± 12.17 %) (p = 0.007). Thus, cine loop stratification using the proposed spatiotemporal CNN model improves the reliability of automated point-of-care echocardiography image interpretation.

Machine Learning Algorithm Detection of Confluent B-Lines.

OBJECTIVE: B-lines are a ring-down artifact of lung ultrasound that arise with increased alveolar water in conditions such as pulmonary edema and infectious pneumonitis. Confluent B-line presence may signify a different level of pathology compared with single B-lines. Existing algorithms aimed at B-line counting do not distinguish between single and confluent B-lines. The objective of this study was to test a machine learning algorithm for confluent B-line identification. METHODS: This study used a subset of 416 clips from 157 subjects, previously acquired in a prospective study enrolling adults with shortness of breath at two academic medical centers, using a hand-held tablet and a 14-zone protocol. After exclusions, random sampling generated a total of 416 clips (146 curvilinear, 150 sector and 120 linear) for review. A group of five experts in point-of-care ultrasound blindly evaluated the clips for presence/absence of confluent B-lines. Ground truth was defined as majority agreement among the experts and used for comparison with the algorithm. RESULTS: Confluent B-lines were present in 206 of 416 clips (49.5%). Sensitivity and specificity of confluent B-line detection by algorithm compared with expert determination were 83% (95% confidence interval [CI]: 0.77-0.88) and 92% (95% CI: 0.88-0.96). Sensitivity and specificity did not statistically differ between transducers. Agreement between algorithm and expert for confluent B-lines measured by unweighted κ was 0.75 (95% CI: 0.69-0.81) for the overall set. CONCLUSION: The confluent B-line detection algorithm had high sensitivity and specificity for detection of confluent B-lines in lung ultrasound point-of-care clips, compared with expert determination.

Two- Versus 8-Zone Lung Ultrasound in Heart Failure: Analysis of a Large Data Set Using a Deep Learning Algorithm.

OBJECTIVE: Scanning protocols for lung ultrasound often include 8 or more lung zones, which may limit real-world clinical use. We sought to compare a 2-zone, anterior-superior thoracic ultrasound protocol for B-line artifact detection with an 8-zone approach in patients with known or suspected heart failure using a deep learning (DL) algorithm. METHODS: Adult patients with suspected heart failure and B-lines on initial lung ultrasound were enrolled in a prospective observational study. Subjects received daily ultrasounds with a hand-held ultrasound system using an 8-zone protocol (right and left anterior/lateral and superior/inferior). A previously published deep learning algorithm that rates severity of B-lines on a 0-4 scale was adapted for use on hand-held ultrasound full video loops. Average severities for 8 and 2 zones were calculated utilizing DL ratings. Bland-Altman plot analyses were used to assess agreement and identify bias between 2- and 8-zone scores for both primary (all patients, 5728 videos, 205 subjects) and subgroup (confirmed diagnosis of heart failure or pulmonary edema, 4464 videos, 147 subjects) analyses. RESULTS: Bland-Altman plot analyses revealed excellent agreement for both primary and subgroup analyses. The absolute difference on the 4-point scale between 8- and 2-zone average scores was not significant for the primary dataset (0.03; 95% CI -0.01 to 0.07) or the subgroup (0.01; 95% CI -0.04 to 0.06). CONCLUSION: Utilization of a 2-zone, anterior-superior thoracic ultrasound protocol provided similar severity information to an 8-zone approach for a dataset of subjects with known or suspected heart failure.

Interobserver Agreement and Correlation of an Automated Algorithm for B-Line Identification and Quantification With Expert Sonologist Review in a Handheld Ultrasound Device.

OBJECTIVES: B-lines are ultrasound artifacts that can be used to detect a variety of pathologic lung conditions. Computer-aided methods to detect and quantify B-lines may standardize quantification and improve diagnosis by novice users. We sought to test the performance of an automated algorithm for the detection and quantification of B-lines in a handheld ultrasound device (HHUD). METHODS: Ultrasound images were prospectively collected on adult emergency department patients with dyspnea. Images from the first 124 patients were used for algorithm development. Clips from 80 unique subjects for testing were randomly selected in a predefined proportion of B-lines (0 B-lines, 1-2 B-lines, 3 or more B-lines) and blindly reviewed by five experts using both a manual and reviewer-adjusted process. Intraclass correlation coefficient (ICC) and weighted kappa were used to measure agreement, while an a priori threshold of an ICC (3,k) of 0.75 and precision of 0.3 were used to define adequate performance. RESULTS: ICC between the algorithm and manual count was 0.84 (95% confidence interval [CI] 0.75-0.90), with a precision of 0.15. ICC between the reviewer-adjusted count and the algorithm count was 0.94 (95% CI 0.90-0.96), and the ICC between the manual and reviewer-adjusted counts was 0.94 (95% CI 0.90-0.96). Weighted kappa was 0.72 (95% CI 0.49-0.95), 0.88 (95% CI 0.74-1), and 0.85 (95% CI 0.89-0.96), respectively. CONCLUSIONS: This study demonstrates a high correlation between point-of-care ultrasound experts and an automated algorithm to identify and quantify B-lines using an HHUD. Future research may incorporate this HHUD in clinical studies in multiple settings and users of varying experience levels.

Automated Lung Ultrasound B-Line Assessment Using a Deep Learning Algorithm.

Shortness of breath is a major reason that patients present to the emergency department (ED) and point-of-care ultrasound (POCUS) has been shown to aid in diagnosis, particularly through evaluation for artifacts known as B-lines. B-line identification and quantification can be a challenging skill for novice ultrasound users, and experienced users could benefit from a more objective measure of quantification. We sought to develop and test a deep learning (DL) algorithm to quantify the assessment of B-lines in lung ultrasound. We utilized ultrasound clips ( n = 400 ) from an existing database of ED patients to provide training and test sets to develop and test the DL algorithm based on deep convolutional neural networks. Interpretations of the images by algorithm were compared to expert human interpretations on binary and severity (a scale of 0-4) classifications. Our model yielded a sensitivity of 93% (95% confidence interval (CI) 81%-98%) and a specificity of 96% (95% CI 84%-99%) for the presence or absence of B-lines compared to expert read, with a kappa of 0.88 (95% CI 0.79-0.97). Model to expert agreement for severity classification yielded a weighted kappa of 0.65 (95% CI 0.56-074). Overall, the DL algorithm performed well and could be integrated into an ultrasound system in order to help diagnose and track B-line severity. The algorithm is better at distinguishing the presence from the absence of B-lines but can also be successfully used to distinguish between B-line severity. Such methods could decrease variability and provide a standardized method for improved diagnosis and outcome.

Enhanced gene expression of systemically administered plasmid DNA in the liver with therapeutic ultrasound and microbubbles.

Ultrasound-mediated delivery (USMD) of novel therapeutic agents in the presence of microbubbles is a potentially safe and effective method for gene therapy offering many desired characteristics, such as low toxicity, potential for repeated treatment, and organ specificity. In this study, we tested the capability of USMD to improve gene expression in mice livers using glycogen storage disease Type Ia as a model disease under systemic administration of naked plasmid DNA. Image-guided therapeutic ultrasound was used in two studies to provide therapeutic ultrasound to mice livers. In the first study, involving wild-type mice, control animals received naked plasmid DNA (pG6Pase 150 µg) via the tail vein, followed by an infusion of microbubbles; the treated animals additionally received therapeutic ultrasound (1 MHz). Following the procedure, the animals were left to recover and were subsequently euthanized after 2 d and liver samples were extracted. Reverse transcription polymerase chain reaction (RT-PCR) assays were performed on the samples to quantify mRNA expression. In addition, Western blot assays of FLAG-tagged glucose-6-phosphatase (G6Pase) were performed to evaluate protein expression. Ultrasound-exposed animals showed a 4-fold increase in G6Pase RNA in the liver, in comparison with control animals. Furthermore, results from Western blot analysis demonstrated a 2-fold increased protein expression in ultrasound-exposed animals after two days ( p < 0.05). A second pilot study was performed with G6Pase knockout mice, and the animals were monitored for correction of hypoglycemia over a period of 3 weeks before tissue analysis. The RT-PCR assays of samples from these animals demonstrated increased G6Pase RNA in the liver following ultrasound treatment. These results demonstrate that USMD can increase gene expression of systemically injected naked pDNA in the liver and also provide insight into the development of realistic approaches that can be translated into clinical practice.

Acoustic accessibility investigation for ultrasound mediated treatment of glycogen storage disease type Ia patients.

Glycogen storage disease type Ia (GSDIa) is caused by an inherited defect in the glucose-6-phosphatase gene. The recent advent of targeted ultrasound-mediated delivery (USMD) of plasmid DNA (pDNA) to the liver in conjunction with microbubbles may provide an alternative treatment option. This study focuses on determining the acoustically accessible liver volume in GSDIa patients using transducer models of various geometries with an image-based geometry-driven approach. Results show that transducers with longer focal lengths and smaller apertures (up to an f/number of 2) are able to access larger liver volumes in GSDIa patients while still being capable of delivering the required ultrasound dose in situ (2.5 MPa peak negative pressure at the focus). With sufficiently large acoustic windows and the ability to use glucose to easily assess efficacy, GSD appears to be a good model for testing USMD as proof of principle as a potential therapy for liver applications in general.

Ultrasound therapy transducers with space-filling non-periodic arrays.

Ultrasound transducers designed for therapeutic purposes such as tissue ablation, histotripsy, or drug delivery require large apertures for adequate spatial localization while providing sufficient power and steerability without the presence of secondary grating lobes. In addition, it is highly preferred to minimize the total number of channels and to maintain simplicity in electrical matching network design. To this end, we propose array designs that are both space-filling and non-periodic in the placement of the elements. Such array designs can be generated using the mathematical concept of non-periodic or aperiodic tiling (tessellation) and can lead to reduced grating lobes while maintaining full surface area coverage to deliver maximum power. For illustration, we designed two 2-D space-filling therapeutic arrays with 128 elements arranged on a spherical shell. One was based on the two-shape Penrose rhombus tiling, and the other was based on a single rectangular shape arranged non-periodically. The steerability performance of these arrays was studied using acoustic field simulations. For comparison, we also studied two other arrays, one with circular elements distributed randomly, and the other a periodic array with square elements. Results showed that the two space-filling non-periodic arrays were able to steer to treat a volume of 16 x 16 x 20 mm while ensuring that the grating lobes were under -10 dB compared with the main lobe. The rectangular non-periodic array was able to generate two and half times higher power than the random circles array. The rectangular array was then fabricated by patterning the array using laser scribing methods and its steerability performance was validated using hydrophone measurements. This work demonstrates that the concept of space-filling aperiodic/non-periodic tiling can be used to generate therapy arrays that are able to provide higher power for the same total transducer area compared with random arrays while maintaining acceptable grating lobe levels.

A novel hands-free carotid ultrasound detects low-flow cardiac output in a swine model of pulseless electrical activity arrest.

OBJECTIVE: To determine if a hands-free, noninvasive Doppler ultrasound device can reliably detect low-flow cardiac output by measuring carotid artery blood flow velocities. We compared the ability of observers to detect carotid artery flow velocity differences between pseudo-pulseless electrical activity (PEA) and true-PEA cardiac arrest. METHODS: Five swine were instrumented with aortic (Ao) and right atrial pressure-transducing catheters. The Doppler ultrasound device was adhered to the neck over the carotid artery. Continuous electrocardiogram, pressure readings, and Doppler signal were recorded. Each swine underwent multiple episodes of fibrillation and resuscitation. Episodes of true-PEA and pseudo-PEA were retrospectively identified from all resuscitation attempts by examination of electrocardiogram and Ao waveforms. The sensitivity and specificity of the device to detect pseudo-PEA was obtained using observers blinded to Ao waveform recordings. RESULTS: There was good interobserver reliability related to identification of pseudo- and true-PEA (κ = 0.873). The observers blinded to Ao waveform recordings agreed on 8 of the 9 episodes of pseudo-PEA, whereas 4 false positives of 26 true-PEA events were reported (sensitivity, 0.89; specificity, 0.85). The Doppler device was able to detect carotid flow velocity over a wide range of Ao blood pressures. CONCLUSIONS: This hands-free, noninvasive Doppler ultrasound device can reliably differentiate pseudo-PEA from true-PEA during resuscitation from cardiac arrest, detecting pressure gradient changes of less than 5 mm Hg through to normotension. This device distinguishes conditions of no cardiac output from low cardiac output and may have applications for use during resuscitation from various etiologies of arrest and shock.

Ultrasound-triggered release of materials entrapped in microbubble-liposome constructs: a tool for targeted drug delivery.

We investigated the preparation of ultrasound-triggered drug delivery system, based on a pendant complex of microbubble coated with liposomes. Biotinylated decafluorobutane microbubbles were coated with biotinylated liposomes via a streptavidin linker. Liposomes were prepared incorporating calcein and thrombin. Based on initial concentration of calcein, over 1 um(3) payload volume per each microbubble-liposome particle was achieved, when 100 nm liposomes were used. Insonation of microbubble-liposome pendants in vitro resulted in the complete destruction of microbubbles and triggered release of a significant fraction of the entrapped material. Treatment with 1MHz ultrasound (5 pulses, 100 ms, 7 MPa peak negative acoustic pressure) resulted in the release of ~30% of entrapped calcein, as estimated by the fluorescence quenching assay. Thrombin release from liposomes complexed with microbubbles (11% of entrapped material) due to ultrasound treatment was estimated by a chromogenic substrate study. Prior to insonation, substrate hydrolysis was at background level. Ultrasound-triggered release of thrombin from the pendant complexes caused an acceleration of blood clotting.

Targeted ultrasound-mediated delivery of nanoparticles: on the development of a new HIFU-based therapy and imaging device.

Ultrasound-mediated delivery (USMD) is an active research topic, as researchers develop applications for therapeutic ultrasound in addition to thermal ablation. In USMD, ultrasound is used in conjunction with microbubbles and drugs, nanoparticles, siRNA, pDNA, stem cells, etc., to facilitate their cellular delivery and uptake using pressure and temperature-mediated mechanisms to bring about a desired therapeutic effect. To investigate the potential of targeted USMD of nanoparticles, pDNA, and stem cells for cardiovascular and other applications, a general-purpose preclinical research tool, therapy imaging probe system (TIPS) was designed. It consists of a wideband annular array, a small-animal acoustic coupler, a motorized positioning system, integrated control software for ultrasound image-guided treatment planning and execution, and triggering electronics that allow ECG and respiration-gated ultrasound exposures. TIPS was then used to enhance delivery of nanoparticles into the murine myocardium and heart vessel walls to demonstrate the feasibility of the technology, pave the way for additional basic research in cardiovascular USMD, and begin to explore the requirements that USMD devices will have to meet to be useful in a clinical setting.

An automated carotid pulse assessment approach using Doppler ultrasound.

During cardiac arrest emergencies, lay rescuers are required to manually check the patient's carotid pulse after the delivery of defibrillation shocks to assess the cardiac resuscitation progress of the patient. As a more automated way of monitoring the resuscitation progress, a new Doppler-ultrasound-based carotid pulse assessment approach is presented in this paper. The method works by analyzing the temporal aperiodicity of Doppler shifts seen in the ultrasound echoes returned from the patient's carotid arteries. As a quantitative investigation with this method, we derived a new measure called the pulselessness indicator to assess whether a carotid pulse is absent based on the given Doppler information. To study the performance of the new carotid pulse checking method, we built a multi-channel CW Doppler prototype device to acquire Doppler data in vivo during cardiac arrest experiments conducted on five different swines and computed pulselessness indicator estimates with these data. Our results indicated that the Doppler-based pulse checking approach has good sensitivity and specificity: it had a pulselessness detection rate greater than 0.9 for a given false alarm rate of 0.05. As a further analysis, the prototype device was applied to other experiments where the swine had suffered cardiac arrest for over five minutes. It showed a consistent assessment performance on the monitoring of the swine's resuscitation progress after defibrillation and chest compressions.

Interlaboratory comparison of ultrasonic backscatter coefficient measurements from 2 to 9 MHz.

OBJECTIVE: As are the attenuation coefficient and sound speed, the backscatter coefficient is a fundamental ultrasonic property that has been used to characterize many tissues. Unfortunately, there is currently far less standardization for the ultrasonic backscatter measurement than for the other two, as evidenced by a previous American Institute of Ultrasound in Medicine (AIUM)-sponsored interlaboratory comparison of ultrasonic backscatter, attenuation, and speed measurements (J Ultrasound Med 1999; 18:615-631). To explore reasons for these disparities, the AIUM Endowment for Education and Research recently supported this second interlaboratory comparison, which extends the upper limit of the frequency range from 7 to 9 MHz. METHODS: Eleven laboratories were provided with standard test objects designed and manufactured at the University of Wisconsin (Madison, WI). Each laboratory was asked to perform ultrasonic measurements of sound speed, attenuation coefficients, and backscatter coefficients. Each laboratory was blinded to the values of the ultrasonic properties of the test objects at the time the measurements were performed. RESULTS: Eight of the 11 laboratories submitted results. The range of variation of absolute magnitude of backscatter coefficient measurements was about 2 orders of magnitude. If the results of 1 outlier laboratory are excluded, then the range is reduced to about 1 order of magnitude. Agreement regarding frequency dependence of backscatter was better than reported in the previous interlaboratory comparison. For example, when scatterers were small compared with the ultrasonic wavelength, experimental frequency-dependent backscatter coefficient data obtained by the participating laboratories were usually consistent with the expected Rayleigh scattering behavior (proportional to frequency to the fourth power). CONCLUSIONS: Greater standardization of backscatter measurement methods is needed. Measurements of frequency dependence of backscatter are more consistent than measurements of absolute magnitude.

3-D finite-element models of human and monkey fingertips to investigate the mechanics of tactile sense.

The biomechanics of skin and underlying tissues plays a fundamental role in the human sense of touch. It governs the mechanics of contact between the skin and an object, the transmission of the mechanical signals through the skin, and their transduction into neural signals by the mechanoreceptors. To better understand the mechanics of touch, it is necessary to establish quantitative relationships between the loads imposed on the skin by an object, the state of stresses/strains at mechanoreceptor locations, and the resulting neural response. Towards this goal, 3-D finite-element models of human and monkey fingertips with realistic external geometries were developed. By computing fingertip model deformations under line loads, it was shown that a multi-layered model was necessary to match previously obtained in vivo data on skin surface displacements. An optimal ratio of elastic moduli of the layers was determined through numerical experiments whose results were matched with empirical data. Numerical values of the elastic moduli of the skin layers were obtained by matching computed results with empirically determined force-displacement relationships for a variety of indentors. Finally, as an example of the relevance of the model to the study of tactile neural response, the multilayered 3-D finite-element model was shown to be able to predict the responses of the slowly adapting type I (SA-I) mechanoreceptors to indentations by complex object shapes.

Robust deconvolution of high-frequency ultrasound images using higher-order spectral analysis and wavelets.

Deconvolution of high-frequency (30-40 MHz) ultrasonic images of human skin was studied in vivo. Separate one-dimensional (1-D) functions for the axial and lateral profiles were first estimated using higher-order spectral methods. Subsequently, deconvolution was implemented us ing a regularized inverse Wiener filtering of the wavelet and scaling coefficients that were obtained after a wavelet decomposition of the RF signals. Deconvolution was first performed in the axial direction, then in the lateral direction. The methods were applied to data obtained from the skin of 16 volunteers using three different transducers. Significant improvements in both the axial and lateral resolutions were obtained in all the cases. Features such as hair follicles in the dermis and fingerprints on the surface of the finger were more clearly displayed in the processed images compared to the original images. The results indicate that the deconvolution method using higher-order spectral methods and wavelet analysis could significantly improve the quality of high-frequency ultrasonic skin images.

Quantitative ultrasonic methods for characterization of skin lesions in vivo.

Quantitative ultrasonic methods were studied for characterizing skin lesions in vivo using contact dermatitis as an example. The parameters studied include skin thickness, echogenicity, attenuation coefficient slope and parameters related to echo statistics (signal-to-noise ratio and shape parameters of Weibull, K and generalized gamma distributions). Data were collected using a high-frequency ultrasound (US) system (center frequency = 33 MHz). To compensate for depth-dependent diffraction effects, correction curves as a function of the distance between the transducer and the tissue were first empirically obtained. Diffraction-corrected quantitative parameters were then compared between healthy and affected skin of volunteers, who underwent patch testing for allergic and irritant contact dermatitis. A significant increase in skin thickness, decrease in echogenicity of the upper dermis and decrease in attenuation coefficient slope were found at the affected sites compared to those of healthy skin. However, no differences in parameters related to the echo statistics of the mid-dermis were found. These results indicate that a combination of quantitative ultrasonic parameters have the potential for extracting information for characterizing skin conditions.

Statistics of envelope of high-frequency ultrasonic backscatter from human skin in vivo.

The statistics of envelope of high-frequency ultrasonic backscatter signals from in vivo normal human dermis and subcutaneous fat were studied. The capability of six probability distributions (Rayleigh, Rician, K, Nakagami, Weibull, and Generalized Gamma) to model empirical envelope data was studied using the Kolmogorov-Smirnov (KS) goodness of fit statistic. The parameters of all the distributions were obtained using the maximum likelihood method. It was found that the Generalized Gamma distribution with two shape parameters provided the best fit among all the distributions in terms of the KS goodness of fit. The K and Weibull distributions also modeled the envelope statistics well. The Rayleigh and Rician distributions provided poorer fits. The parameters of the Generalized Gamma distribution, however, showed a larger variability than those of the other distributions. The intersubject variability in the estimated parameters of all the distributions was found to be comparable to the intrasubject variability. Fat was seen to exhibit significantly more pre-Rayleigh behavior compared to the dermis. The parameters of the Generalized Gamma distribution also showed significant differences between the dermis at the forearm and fingertip regions.