High contrast power Doppler imaging in side-viewing intravascular ultrasound imaging via angular compounding

Abstract

The ability to assess likelihood of plaque rupture can determine the course of treatment in coronary artery disease. One indicator of plaque vulnerability is the development of blood vessels within the plaque, or intraplaque neovascularization. In order to visualize these vessels with increased sensitivity in the cardiac catheterization lab, a new approach for imaging blood flow in small vessels using side-viewing intravascular ultrasound (IVUS) is proposed. This approach based on compounding adjacent angular acquisitions was evaluated in tissue mimicking phantoms and ex vivo vessels. In phantom studies, the Doppler CNR increased from 3.3 ± 1.0 to 13 ± 2.6 (conventional clutter filtering) and from 1.9 ± 0.15 to 7.5 ± 1.1 (SVD filtering) as a result of applying angular compounding. When imaging flow at a rate of 5.6 mm/s in 200 μm tubes adjacent to the lumen of ex vivo porcine arteries, the Doppler CNR increased from 5.3 ± 0.95 to 7.2 ± 1.3 (conventional filtering) and from 23 ± 3.3 to 32 ± 6.7 (SVD filtering). Applying these strategies could allow increased sensitivity to slow flow in side-viewing intravascular ultrasound imaging.

Introduction

Coronary artery disease is responsible for 400,000 deaths annually in the U.S. [1], and the World Health Organization ranks it as the number one cause of mortality globally (13.2% of deaths as of 2012) [2]. The rupture of coronary atherosclerotic plaques can lead to major adverse cardiac events (MACE) such as myocardial infarction. The ability to stratify risk of rupture for a given lesion could reduce rates of MACE. However, because many patients are asymptomatic prior to rupture (i.e. stable coronary artery disease), new methods of characterizing plaque vulnerability and risk of MACE are needed [3] [4]. As plaques develop, microvessels called vasa vasorum invade the plaque from the adventitial layer of the arterial wall [3]. Vasa vasorum density in the plaque is correlated with stenosis severity in humans [5] and with plaque vulnerability in animal models of atherosclerosis [6] [7]. The ability to stratify risk of rupture in patients with stable disease could determine which lesions would benefit from intervention (e.g. stenting).

Currently, interventionalists have several tools for assessing lesion severity. Imaging techniques for evaluating local plaque structure include catheter angiography [8], magnetic resonance and computed tomography angiography [9], optical coherence tomography (OCT) [10] [11], and intravascular ultrasound. While each modality has advantages and disdvantages, intravascular ultrasound (IVUS) provides relatively high penetration depth imaging of the vessel wall, plaque, and fibrous cap without the use of ionizing radiation. Furthermore, analysis of acquired radiofrequency IVUS data has enabled plaque characterization, as integrated backscatter IVUS and virtual histology IVUS have demonstrated the ability to predict plaques causing acute coronary syndrome and to accurately assess plaque composition, respectively [12] [13]. While there are reports of imaging vasa vasorum with IVUS, current techniques for flow imaging with IVUS have difficulty identifying slow blood flow in the small vessels of the vasa vasorum (60 μm to several hundred μm), particularly in the smallest vessels with the slowest flow [14] [15] [16] [17].

Recently, several groups have presented new technical developments to improve IVUS imaging capabilities. Researchers have demonstrated multi-frequency transducers to provide high resolution imaging of near structures without loss of penetration depth [18] [19] [20]. In addition, other groups have presented new developments in IVUS signal processing techniques, including demonstration of improved lateral resolution due to sub-aperture beamforming [21] or deconvolution [22]. In order to image intraplaque neovascularization, several researchers have demonstrated contrast agent-specific IVUS imaging [23] [24] [25]. Using a subharmonic approach for contrast-specific imaging, Goertz et al. demonstrated increased contrast-to-tissue ratio (CTR) by up to 30 dB relative to non-linear imaging in atherosclerotic rabbit vessels [26]. Shekhar et al. demonstrated further increase in CTR relative to subharmonic imaging by combining subharmonic and ultraharmonic echoes [27]. Alternatively, other groups have recently proposed promising approaches for intravascular ultrasound imaging including intravascular spectroscopic photoacoustic imaging [28] and intravascular elastography (palpography) [29] [30].

While the ability to characterize plaques based on intraplaque neovascularization by imaging blood flow in vasa vasorum could guide clinical decision-making [31] [32], current blood flow imaging with side-viewing IVUS is limited to the largest vessels of the vasa vasorum due to the difficulty of separating signals arising from slow blood flow in small vessels from surrounding higher amplitude tissue signals [16] [33] [34]. In side-viewing IVUS imaging, as the single element transducer rotates through 360°, the ultrasound beam intersects with the vasa vasorum at a variety of angles both in and out of the imaging plane. In clinical IVUS systems operating at 1800 revolutions per minute (30 per second), the slow time sampling rate at a given location is 1/30 s = 33.3 ms, which prevents estimation of blood flow velocity or direction using conventional Doppler techniques due to the low pulse repetition frequency (PRF). If the flow velocity is higher than the Nyquist limit for slow time sampling, as is the case in this work (i.e. >1mm/s flow) and other rotational IVUS applications [35] [36] [37] [38], then none of the same scatterers are sampled from one revolution to the next. Out-of-plane flow also results in new scatterers entering the imaging plane on each successive measurement. However, by filtering stationary or slow-moving structures observed in subsequent rotations, images can be formed that localize blood flow, analogous to power Doppler images in ultrasound imaging with array transducers [15] [39] [40]. As seen in various IVUS applications [35] [36] [37] [38], flow above the system’s Nyquist limit and with velocity components in the elevation direction can be visualized without reporting velocities, which can prove useful for some clinical scenarios.

In this article, we present approaches for increasing sensitivity to blood flow in mechanically-steered IVUS imaging. By compounding wall-filtered data from adjacent angles, Doppler contrast-to-noise ratio can be increased. In array-based ultrasound, compounding is commonly used to improve the point spread function and thus increase contrast and spatial resolution while also increasing SNR [41] [42]. In this work, angular compounding to increase flow sensitivity in mechanically-steered IVUS data is demonstrated for the first time, using both conventional and SVD filtering to suppress clutter from stationary tissue. The sensitivities of these processing approaches are evaluated in rotational IVUS studies in tissue-mimicking phantoms with introduced displacements to imitate physiological motion and in ex vivo porcine arteries.

Materials and Methods

In this work, angular compounding was applied to rotational IVUS to increase SNR and thus sensitivity to slow flow in small vessels. In order to determine the angular extent over which angular compounding should be applied, the angular correlation between adjacent acquisitions was experimentally assessed in a tissue-mimicking phantom to ensure only highly-correlated signals were compounded to increase signal-to-noise ratio, and in turn increase the ability to separate signals arising from flow from those arising from stationary or slow-moving tissue. Tissue echoes were filtered using either conventional (Butterworth) or singular value decomposition (SVD) filtering. SVD filtering has previously been applied in array-based ultrasound imaging to separate tissue motion from blood flow [43], as the first few high-power singular values typically correspond to highly correlated tissue signals, while the other singular values typically contain both flow signals with lower slow time correlation and uncorrelated random noise [43]. For SVD filtering, high power singular values with high spatiotemporal coherence above the point of inflection of the decaying curve are removed [44] [45]. As the ensemble size allowed in rotational IVUS due to lower frame rate (30 Hz as opposed to >1kHz) is relatively small (30 frames result in 30 singular values), the turning point of the singular values was observed to be relatively stationary, meaning a simplifying assumption of a removal of a constant number of singular values could be used for this work.

Transducer and data acquisition system

Radiofrequency (RF) echo data were acquired using a custom laboratory IVUS system with a 1.1 × 0.55 mm, side-viewing 20 MHz transducer [34]. System characteristics are described in Table 1, and the prototype imaging setup is shown in Figure 1. The transducer was mounted on a stepper motor (step size 0.225°) controlled by a microcontroller (Arduino UNO, Torino, Italy). At each angular position, the transducer was excited by a negative impulse with a peak amplitude of −227 volts (Panametrics 5900PR, Waltham, MA, USA), then radiofrequency (RF) data were digitized at a sampling rate of 100 MHz (PDA14, Signatec, Corona, CA) using a custom program (LabVIEW, National Instruments Corp., Austin, TX). During data acquisition, the transducer was moved relative to the phantom using a 3-axis motion stage (Newport XPS, Irvine, CA) to simulate physiological motion. After each rotation, the transducer was translated in the elevation direction by the motion stage to provide 3D data. The slow time period between subsequent acquisitions at the same spatial position was ~33 seconds to allow a full rotation.

Table 1:

Parameters for rotational IVUS setup

Aperture Size	0.55 × 1.1 mm
On-axis −6 dB Pressure range	0–13 mm (max 7mm)
Center Frequency	20 MHz
Axial Resolution	0.15 mm
Angular Step	0.225°
Elevational Step	1 mm

Figure 1: Acquisition system for rotational imaging.

The transducer (yellow) is mounted to a catheter (black), which is rotated by a stepper motor to provide side-viewing acquisitions for 360° rotation through the imaging plane (shown by curved blue arrows). After each revolution, a motion stage steps the transducer by 1 mm in the elevation direction (shown by the orange arrow), which allows 3D imaging. To reflect physiological conditions, the transducer was translated in the imaging plane while acquiring data, as indicated by the red arrow.

Motion correction

Due to the low frame rate, motion correction is needed between acquisitions to align acquired data in the presence of simulated tissue motion. Several motion correction strategies have previously been employed in IVUS to correct for physiological motion [46] [47] [48] [49] [50]. In this work, acquired 2D frames were aligned using a local block matching method similar to the method previously described in [51]. Briefly, 2D block matching using relatively large kernels (6.8 mm × 10 angles) and a search region of 7 mm × 20 angles was used to determine the delays maximizing cross-correlation between all adjacent frames. False matches were reduced by limiting the search region to velocities less than 20 mm/s. This excludes unrealistic tissue velocities, and the threshold can be set based on the imaged region and known tissue velocities. After curve fitting to prevent discontinuous, non-physical corrections [52] [53] [54] [55], for each frame, entire lines of acquired RF data were aligned with respect to the central frame (e.g. frame 15 for an ensemble size of 30).

Power Doppler processing

The processing approaches are illustrated in Figure 2. Before the data were clutter filtered in the slow time domain, acquired RF data were bandpass filtered with a fourth order Butterworth filter (14–28MHz) to remove noise and upsampled to 400 MHz, then motion correction was performed as described previously. Next, one of four different post-processing strategies was applied to form the power Doppler images (Fig. 2): A) Butterworth high pass filtering without angular compounding, B) SVD filtering without angular compounding, C) Butterworth high pass filtering with angular compounding, and D) SVD filtering with angular compounding. An ensemble size of 30 frames was used. All processing was performed in Matlab (The Mathworks, Natick, MA, USA).

Fig. 2. Processing schematic for motion-corrected power Doppler images.

Four types of processing were compared: (A) Images with Butterworth filtering, (B) SVD filtering, (C) Butterworth filtering with compounding, and (D) SVD filtering with compounding.

For Butterworth filtering schemes, motion-corrected RF data were filtered in slow time to remove slow-moving tissue echoes [56], and the complex envelope was estimated. Frequency cutoffs equivalent to 3 and 12 Hz in a rotational system acquiring at 30 revolutions per second were used, which correspond to 0.12 to 0.46 mm/s along the beam. After varying parameters, these cutoffs were observed as consistently effective for removing both tissue and noise. As an initial comparison with the motion-corrected Butterworth filter case, RF data before motion correction were Butterworth filtered in slow time.

For SVD filtering, each angular location of the 2D imaging region was filtered individually to ensure high local spatial correlation. This process is illustrated in Figure 3. Briefly, a 3D ensemble of data was collected (axial depth × circumferential angle × number of frames), as shown in Figure 3-A. For each angle, a 2D matrix (axial sample × number of frames), depicted in Figure 3-B, was decomposed using SVD into a product of three matrices, generally referred to as U, Σ, and V, where columns of U and V are orthonormal bases (Figure 3-C). The diagonal of the central diagonal matrix, Σ, was then filtered to remove singular values corresponding to tissue signal (Fig. 3-D). Though plane wave imaging has many frames which leads to an increase in cumulative noise [44], the frame rate of rotational IVUS is much lower, similar to optical microangiography. Yousefi et al found that the clutter space was represented in just their first two singular values and their removal allowed adequate filtering [57]. For this study with an ensemble size of 30, the first two singular values were observed to correspond to tissue signals with high spatiotemporal correlation and thus were discarded.

Figure 3: SVD filtering of individual A-lines.

A) An ensemble of data is acquired. B) For each individual angle, C) data are decomposed, and D) the singular values in Σ corresponding to tissue signal are removed. L1 denotes the number of fast time samples, while L2 denotes the number of slow time samples which is 30 in the present study.

For cases (C) and (D) with compounding, the SVD-filtered or Butterworth-filtered, envelope-detected data were compounded over the number of angular acquisitions for which the correlation coefficient exceeded 0.5 in preliminary experiments in tissue speckle-producing phantoms as described in the next section. The same filter parameters were utilized in all experiments.

Determination of compounding angle

Sub-aperture processing in rotational, mechanically-steered IVUS imaging has previously demonstrated increased SNR and contrast relative to single acquisition processing without summation [21]. In this work, envelope-detected data rather than RF data or complex analytic data were compounded, similar to methods with array transducers [58][59][60][61][62]. As the angle between acquisitions increases, signals become increasingly decorrelated, thus summing contributions over too broad an angular extent would result in decreased SNR. Conversely, compounding over an angular extent that is too narrow fails to realize the maximum SNR improvement.

Angular correlation in rotational IVUS was investigated using two custom tissue mimicking phantoms: one containing only diffuse targets (α=0.5 dB/cm/MHz, SNR_speckle = 1.68 ± 0.06) [63], and a second phantom containing a 140 μm wire embedded in diffuse scatterers (α=0.5 dB/cm/MHz). These phantoms allow assessment of angular correlation for both diffuse and coherent targets. 50 independent sets of RF echo data were acquired from this phantom using the described setup.

RF data were bandpass filtered with a 4^th order Butterworth filter (14–28 MHz) and upsampled to 400 MHz. In order to assess angular correlation, the normalized correlation coefficient was computed according to the following equation [64]:

$$\rho =\frac{\sum _{n=1}^N\left[s_{\theta }(n)-{\mu }_{\theta }\right]{\left[s_{\theta +\Delta \theta }(n)-{\mu }_{\theta +\Delta \theta }\right]}^{*}}{\sqrt{\sum _{n=1}^N{\left[s_{\theta }(n)-{\mu }_{\theta }\right]}^2\sum _{n=1}^N{\left[s_{\theta }+\Delta \theta (n)-{\mu }_{\theta }+\Delta \theta \right]}^2}}$$ (Equation 1)

Where n is the time index of the RF data along the beam, i.e. in the axial direction, N is the number of samples per segment, s is the acquired RF signal at a single angular location θ, θ + Δθ indicates the acquired RF signal at a subsequent angle, and μ is the mean of signal s. The correlation coefficient as a function of angle was determined for 1.5 mm segments of acquired RF lines spaced every 2 mm centered at depths from 2–30 mm. For each depth, the angular extent for compounding was determined based on the maximum angular extent for which ρ ≥ 0.5 based on 50 acquisitions of windowed RF data.

Experimental studies: tissue-mimicking phantoms

The performance of flow imaging schemes was assessed using a tissue-mimicking phantom (α=0.5 dB/cm/MHz) with an embedded cellulose tube (inner diameter = 0.2 mm, Spectrum Labs, Rancho Dominguez, CA, USA) located at a depth of 7 mm from the transducer. Though flow in vasa vasorum can exist in multiple dimensions relative to the transducer, the elevation component of flow is the dominant direction [65] (as well as the most challenging to separate) and was investigated with our experimental setups. While ultimately the goal is to image blood flow with high sensitivity without the use of exogenous contrast agents, microbubbles were injected into the tube such that the maximum signal amplitude was 15 dB lower than the maximum artery wall amplitude in order to have a scattering source with controllable amplitude to produce the desired intensity ratio. Microbubbles were synthesized in house using a mixture of 1,2-distearoyl-sn-glycero-3-phosphocholine (DSPC) and 1,2-distearoyl-sn-glycero-3-phosphoethanolamine-N-methoxy (polyethylene–glycol)-2000 (DSPE-PEG2000; Avanti Polar Lipids, Alabaster, AL, USA) in a 9:1 molar ratio was dissolved in phosphate-buffered saline (80% v/v), propylene glycol (15% v/v), and glycerol (5% v/v). Microbubbles with lipid shells and perfluorobutane core were formed by filling headspace of vials containing lipid solutions with perfluorobutane gas (Fluoromed, Round Rock TX) using a custom gas exchanging apparatus, followed by mechanical agitation for 45 seconds (Vialmix, Lantheus Medical Imaging, Billerica, MA). 10⁸ microbubbles / mL were infused at a rate of 5.6 mm/s (PHD2000, Harvard Apparatus, Holliston, MA) based on published values for flow velocities in the microvasculature [33] [34].

To reflect the reported scale of arterial wall displacement [66] [67] [68] with a commercial IVUS system operating at 30 revolutions per second, constant linear motion (0.66 mm per 30 transducer rotations) of the transducer relative to the phantom was introduced in the imaging plane (Fig. 1) using the motion stage. The presence of relative motion between the IVUS transducer and the artery is a significant challenge that is partially addressed with the block-matching motion correction processing.

For each processing approach, the effectiveness of flow imaging was quantified by computing the Doppler CNR [43]:

$$CN{R}_{\text{Doppler}}=\frac{\overline{P{W}_{\text{flow}}}-\overline{P{W}_{\text{tissue}}}}{std\left(P{W}_{\text{tissue}}\right)}$$ (Equation 2)

Where $\overline{PW}$ is the mean power Doppler signal in the specified region (ensemble size of 30 frames). Significance of observed differences in Doppler CNR with and without compounding was assessed using paired t-tests with an alpha level of 0.05.

Simulation studies: Arteries

In order to assess performance of the proposed technique for motion in different directions, a 3D artery with coherent regions was simulated using Field II [69]. Either in-plane or out-of-plane tissue motion was introduced at a rate of 0.66 mm per 30 transducer rotations to reflect the amplitude of physiological motion [70] [71]. A 0.2 mm diameter region of flow was created by moving low amplitude scatterers. Ensembles of 30 frames were acquired. For out-of-plane motion, tissue motion with the same introduced velocity was added in the elevational direction. The maximum amplitude of scatterers in the region of flow was 15 dB lower than the amplitude of the simulated artery wall. Four realizations were created with different random amplitudes and position of scatterers. Amplitudes of scatters were assigned from a Gaussian distribution offset to be greater than zero, and positions of scatters were randomly generated and retained within the artery wall. RF data were processed as described (see section “Power Doppler processing”). For each processing approach, the effectiveness of flow imaging was quantified by computing the Doppler CNR

Experimental studies: Ex vivo arteries

An excised porcine carotid artery (Sierra Medical, Whittier, CA) was fixed in a custom frame to maintain tension, and 200 μm-diameter cellulose tubing was positioned adjacent to the inner wall of the lumen parallel to the long axis of the vessel to mimic a microvessel. The distance from the transducer to the vessel wall and adjacent cellulose tubing was approximately 4 mm. Flow of microbubbles was introduced inside the tubing at 5.6 mm/s as in tissue-mimicking phantom experiments. 3D data sets were acquired with 13 slices in the elevation direction at a step size of 1 mm. RF data were processed independently slice-by-slice as described (see section “Power Doppler processing”). Doppler CNR was quantified for each case according to Equation 2. Additionally, 3D images were created in ImageJ by creating projections of the 3D data displaying the brightest point.

Results

Determination of compounding angle

Representative correlation as a function of angle is shown in Figure 4 for acquired radiofrequency data, where multiple acquisitions (n = 50) were achieved by stepping in the elevation direction to image different cross sections (full rotations of the transducer) of the coherent and diffuse targets. This relationship is analogous to linear decorrelation described by the van Cittert-Zernike theorem [72]. ρ remains close to 1 for the coherent target case as the transducer rotates through a small angular extent. However, after approximately six degrees of rotation, the correlation coefficient begins to decrease. Meanwhile, the correlation coefficients for the diffuse case decrease more rapidly.

Figure 4: Experimentally-measured spatial correlation as a function of angle at a depth of 5 mm for a coherent target (blue) and a diffuse target (orange) in attenuating media.

These results are used to determine the angular extent over which adjacent acquisitions are compounded. The solid line is themean of fifty measurements and the shaded region represents one standard deviation.

For each axial window, the angle at which the correlation coefficient decreases below 0.5 was determined and used to establish the angular extent of acquisitions to compound at a given depth. The angular extent for compounding as a function of depth is shown in Figure 5. The plot presents proposed values based on analysis rather than measurements with variability, so error bars are not included in this visualization of compounding thresholds. Again, these angular extents were used in compounding in subsequent experiments.

Figure 5: Proposed angular extent for compounding as a function of scan depth.

Acquired data are compounded for all angles over which ρ ≥ 0.5 based on 50 experimental measurements from a diffuse target. As depth increases, the angular extent of compounding initially increases, reaching a maximum at approximately 8 mm before decreasing, as deeper targets decorrelate more rapidly.

Experimental studies: Tissue-mimicking phantoms

Power Doppler images acquired with the 200 μm tube in the tissue-mimicking phantom with introduced motion in the imaging plane are shown for Butterworth filtering both with and without motion correction in Figure 6A-B. Scan converted images are presented with flow (5 dB dynamic range) overlaid on conventional grayscale IVUS images (35 dB dynamic range). For Butterworth filtering without motion correction, flow is difficult to visualize (Fig. 6A), while a region of flow can be visualized more easily in the case with motion correction (Fig. 6B). All further presented results include motion correction. Power Doppler images acquired in the tissue-mimicking phantom with Butterworth and SVD filtering with and without compounding are presented in Figure 7. Though the flow is not apparent in B-mode images, it is visible with Butterworth and SVD filters (Fig. 7A-B), and the area of false flow artifact outside of the region of flow is reduced in the compounded images (Fig. 7C-D), as the amplitude outside of the region of flow appears visibly lower. Overall, the tissue signal of lowest amplitude is seen in the case of Butterworth filtering with compounding, and large smooth areas of flow are seen for Butterworth filtering with compounding as well as SVD filtering with compounding.

Figure 6: Motion correction enables higher sensitivity flow imaging.

Power Doppler images were formed of a 200 μm diameter cellulose tube in a tissue-mimicking phantom via A) Butterworth filtering without motion correction and B) Butterworth filtering with motion correction. Flow was introduced at 5.6 mm/s. Power Doppler data (red) are displayed with 5 dB dynamic range, and B-mode data (gray) are displayed with 35 dB dynamic range. Scale bar indicates 1 mm. The experimental setup is shown in (C), with the grey box indicating the region of interest in (A) and (B), the red dot representing the region of flow, and the black circle representing the transducer. The entire setup is submersed in a water tank (blue). The arrows in A denote residual tissue signal above 5 dB.

Figure 7. Motion-corrected power Doppler images of flow in a 200 μm tube embedded in a tissue-mimicking phantom for images formed via A) Butterworth, B) SVD, C) Butterworth with compounding, and D) SVD with compounding.

Power Doppler data are displayed with 10 dB dynamic range.

In Figure 8, motion-corrected power Doppler data is displayed with 5 dB dynamic range overlaid onto B-mode data. Qualitatively, the compounded images show larger regions of flow above at the display threshold. Quantitative results indicate increased Doppler CNR due to compounding, with values of 3.3 ± 1.0 (Butterworth, no compounding), 13 ± 2.6 (Butterworth, with compounding), 1.9 ± 0.15 (SVD, no compounding), and 7.5 ± 1.1 (SVD, with compounding), as illustrated in Figure 9. Multiple realizations were achieved by imaging four unique phantom cross sections. Paired t-tests indicate significant increases in Doppler CNR due to compounding (p=0.0099 for Butterworth and p=0.014 for SVD).

Figure 8. Motion-corrected power Doppler images of flow in a 200 μm tube (red) overlaid on B-mode data (grayscale) of the tissue-mimicking phantom surrounding the tube for images formed via A) Butterworth, B) SVD, C) Butterworth with compounding, and D) SVD with compounding.

Power Doppler data are displayed with 5 dB dynamic range, and B-mode data are displayed with 35 dB dynamic range. Scale bar indicates 1 mm.

Figure 9: Doppler CNR for Butterworth and SVD filtering without (blue) and with compounding (orange) for flow in a 200 μm tube in a tissue-mimicking phantom.

N=4, and the asterisks denote statistical significance using paired t-tests with alpha level of 0.05.

Simulation studies: Arteries

An illustrative result of simulated imaging with in-plane and out-of-plane tissue motion is shown in Figure 10, with B-mode data displayed in grayscale and power Doppler data displayed in red. Flow can be seen in the region of introduced flow, not in the wall of the moving vessel. For introduced motion in the imaging plane, Doppler CNR values were 0.7 ± 0.3 (Butterworth, no compounding), 0.64 ± 0.4 (Butterworth, with compounding), 17.7 ± 11.6 (SVD, no compounding), and 26.2 ± 14 (SVD, with compounding). For out-of-plane tissue motion, Doppler CNR were 1.01 ± 0.3 (Butterworth, no compounding), 1.04 ± 0.3 (Butterworth, with compounding), 2.31 ± 1.2 (SVD, no compounding), and 3.32 ± 1.3 (SVD, with compounding).

Figure 10. Motion-corrected power Doppler data (red) overlaid on B-mode data (grayscale) for a 200 μm region of flow adjacent to the inner surface of a simulated vessel for in-plane (a) and out-of-plane (b) motion.

Power Doppler data are displayed with 5 dB dynamic range, and B-mode is displayed with 35 dB dynamic range. The primary scale bar indicates 2 mm, and the scale bar for the inset indicates 1 mm.

Experimental studies: Ex vivo arteries

Images of flow in a 200 μm tube positioned on the inner surface of an ex vivo porcine artery are shown in Figure 11. Note that the tube wall produced echoes which can be seen in B-mode images. In looking at the wall of ex vivo vessel, the maximum backscattered amplitude from the vessel wall in B-mode is 15 dB greater than the maximum amplitude from flow within the tube, and the mean backscattered amplitude is more than 6 dB higher in the wall of the ex vivo vessel than in the region of flow.Fig. 11-A and 11-C contain residual vessel wall signal visible in images formed with Butterworth filtering. Images formed with SVD filtering do not have visible vessel wall signal above 5 dB. For both Butterworth and SVD filtering, compounded images have a larger, more contiguous flow area above the display threshold. Additionally, compounding resulted in increased visibility of flow in the tube. Doppler CNR values were 5.3 ± 0.95 (Butterworth, no compounding), 7.2 ± 1.3 (Butterworth, with compounding), 23 ± 3.3 (SVD, no compounding), 32 ± 6.7 (SVD, with compounding), as presented in Figure 12. Multiple realizations were achieved by imaging four different artery cross sections. Paired t-tests indicate significant increases in Doppler CNR due to compounding (p=4.8 × 10⁻⁵ for Butterworth and p=0.0036 for SVD). Finally, 3D views of the porcine artery with overlaid flow from SVD filtering with compounding are shown in Figure 13. The grayscale image reveals the morphology of the vessel wall, while the flow within the tube is observed in the overlaid flow image.

Figure 11. Motion-corrected power Doppler data (red) overlaid on B-mode data (grayscale) for a 200 μm tube positioned adjacent to the inner surface of a 7 mm ex vivo porcine vessel for images formed via A) Butterworth, B) SVD, C) Butterworth with compounding, and D) SVD with compounding.

Power Doppler data are displayed with 5 dB dynamic range, and B-mode is displayed with 35 dB dynamic range. The arrows denote residual tissue signal above 5 dB. The insets amplify the region of interest. The primary scale bar indicates 2 mm, and the scale bar for the inset indicates 1 mm.

Figure 12: Doppler CNR for Butterworth and SVD filtering without (blue) and with compounding (orange) for flow in a 200 μm tube positioned on the inner surface of a 7 mm ex vivo porcine artery.

N=4, and the asterisks denote statistical significance using paired t-tests with alpha level of 0.05.

Figure 13: Motion-corrected power Doppler data (red) overlaid on B-mode data (grayscale) for a 200 μm tube positioned adjacent to the inner surface of a 7 mm ex vivo porcine vessel for images formed via SVD filtering with angular compounding.

Each view represents the brightest point along the ray line from a different viewing angle.

Discussion

Tissue echo amplitude is typically far greater than the backscattered signal from blood [73]. For this reason, as well as the challenge of cancelling tissue echoes, the signal-to-noise ratio of blood flow imaging is inherently low. Thus, even a modest increase in post-clutter filter SNR can produce significant improvements in the sensitivity of flow imaging. By compounding adjacent angular acquisitions with ρ ≥ 0.5 in preliminary studies in speckle-producing phantoms, a statistically significant increase in Doppler CNR was observed for each filter and imaging scenario.

Transducer and data acquisition system

While this work used a relatively large, lower frequency transducer to demonstrate feasibility of the proposed processing approaches, a smaller, higher frequency transducer (≥35 MHz) would be required to image coronary arteries. Use of a higher frequency transducer would also result in increased scattering from red blood cells (and from stationary targets) as well as improved resolution. The 20 MHz transducer used in this work could resolve 200 μm targets; however, vessels with diameters less than 100 μm would not be resolved. Depending on the order and location of the vasa vasorum [74], lumen size varies significantly, wherein intraplaque neovascularization represents some of the smallest microvessels with sizes ranging from 60 to several hundred μm. As detailed in Goertz et al., small vessels would appear larger than their true size [26] [71], and adjacent vessels may appear as a single vessel. If vessels are combined, metrics that use vasa vasorum density in number per unit area underestimate the true density, impairing their ability to predict plaque rupture. Use of a higher frequency transducer would allow investigation with smaller vessels. If each individual vessel cannot be resolved, the ability to separate flow from tissue may still be useful for estimating total flow.

Additionally, dual-frequency IVUS transducers have been shown to allow high resolution B-mode images while allowing harmonic imaging [75] [76]. With minor increases in system complexity, we would expect significant improvements in resolution near the transducer where it is most important for our application as well as the characterization of thin cap fibroatheromas.

Motion correction

As illustrated in Figure 6, motion correction greatly improves the sensitivity to slow blood flow in IVUS even in the presence of slow flow velocity and small displacements between acquisitions. In the presented studies in phantoms, both motion-corrected Butterworth and SVD filters demonstrate the ability to separate slow flow from a slow-moving tissue-mimicking phantom. When the tissue echoes consisted of diffuse targets only in the phantom case, the Butterworth filter performed similarly to the SVD filter, though compounding increased Doppler CNR by a greater amount relative to compounding the SVD-filtered data. For each filter type, Doppler CNR values increased by approximately a factor of four due to compounding.

Determination of compounding angle

In assessing angular correlation for a diffuse target, the correlation coefficient decreased below 0.5 within the first 2°, similar to the linear decrease in ρ with linear translation case as described by the van Cittert-Zernike theorem in [72], while for a coherent target, ρ remained ≥ 0.7 until an angle of approximately 6° was reached before decaying due to the departure of the target from the beam as well as different paths through the field of graphite scatterers. The determined angular extent for compounding initially increased with depth, then decreased due to the variation of beam width in depth [34]. Consequently, by compounding over a limited angular extent as a function of depth, only more highly correlated signals were compounded [77]. With the current pre-selected threshold of ρ ≥ 0.5, the compounding angular extent varied from three to twenty-three angular positions (0.50° to 5.0° with step size 0.225°) as a function of depth (Fig. 4). Previously, others have beamformed IVUS data using constant windows of one, three, or seven (0 to 5.4° with a setup with step size 0.9°) in order to increase B-mode sensitivity or resolution [21]. The increased SNR due to compounding correlated echoes results in an increase in sensitivity to slow-moving flow due to suppression of electronic noise, which allows low amplitude, high frequency flow signals to be detected above the noise. In the future, the correlation coefficient and corresponding angular extent for compounding could be varied based on incoming data to dynamically increase Doppler sensitivity. Finally, as described in Figure 5, the variation in beam width with depth influences the angular extent of compounding. A wider beam would yield higher correlation over a greater angular extent. However, spatial resolution would be lost. In the case of a smaller, higher frequency transducer, this might necessitate a smaller angular step size, i.e. a higher PRF for a continuously rotating system.

Experimental studies: tissue-mimicking phantoms

The presented flow experiments used tubing with an inner diameter of 0.2 mm, whereas secondary vasa vasorum diameters can be <0.1 mm [78] [17]. The sensitivity limits for the presented techniques as a function of flow volume, blood velocity, and tissue motion will need to be determined, and a higher acquisition rate system will be required to realize the presented benefits for in vivo imaging. In addition, studies need to be performed without exogenous contrast agents to assess Doppler CNR as a function of vessel diameter, flow rate, and tissue motion rate using only blood or an acoustic blood-mimicking fluid to determine the limitations of slow flow imaging with side-viewing IVUS.

If sensitivity is not sufficient for imaging blood flow in the smallest vessels of the vasa vasorum without contrast agents, the presented technique might be utilized in combination with microbubbles, with the potential benefit of not requiring a contrast-specific transducer or system. Further, great progress has been made in the development of contrast-enhanced ultrasound imaging systems [79] [80] [81], and the fusion of these specialized systems with our proposed processing would likely be synergistic. The approach presented here might also be combined with approaches for non-linear imaging of microbubbles [82] [83] using commercial systems.

In the future, it might be possible to implement the demonstrated processing approaches for real time imaging. Rapid motion correction has been demonstrated [84], and the described filters are already utilized in real time ultrasound imaging [43] [85]. In a future IVUS imaging system, the demonstrated approaches could be translated to an IVUS system operating with a PRF of 40 kHz rotating at 1800 RPM. These parameters allow data acquisition at 30 frames per second with 1320 lines per rotation, comparable to the 1600 that were used in this study. Tissue speeds can range from stationary to several centimeters per second, while blood generally moves between approximately 5 mm/s and 20 cm/s within the vasa vasorum [86] [55]. Therefore, depending upon the local physiology, filter parameters may need to be adjusted to adaptively optimize image quality [44] [87].

While increasing the ensemble size could increase the Doppler CNR [88], a larger ensemble at a given frame rate (i.e. 30 revolutions per second) would require a longer acquisition time which may include a period of high velocity tissue motion that could degrade the image quality. Alternatively, to increase the ensemble length without increasing the total acquisition time, a higher scanning frame rate could also be achieved using an electronically-steered IVUS array. However, the angular step size would be limited in a side-viewing circular array relative to rotational IVUS, decreasing the number of acquisitions that could be compounded.

Simulation studies: Arteries

Both in-plane and out-of-plane motion were modeled in simulations. The 2D block matching motion correction strategy was able to compensate for in-plane motion, producing positive Doppler CNR values. While most of the tissue was suppressed for the simulation with out-of-plane tissue motion at the same velocity, Doppler CNR values were lower than the corresponding values for in-plane tissue motion, and flow artefact remained in some regions of the artery wall in the displayed image. It will be necessary to investigate the amplitudes of physiological motion for the targeted applications as they will alter the system design in terms of aperture and pulse repetition frequency. Future work with actual physiological motion both in-plane and out-of-plane with nonlinear velocity profiles in animals will be helpful for determining the effectiveness of the proposed processing. Improved motion compensation may be needed, or the results could suggest the use of an IVUS array to allow correction of out-of-plane motion.

Experimental studies: Ex vivo arteries

In imaging flow adjacent to the ex vivo vessel, the filtered signal allowed separation of the region of flow from the vessel even though the echo intensity of the vessel was far greater than echoes within the tube. In clinical settings, there would also be flow within the lumen, which might increase the difficulty of detecting flow in microvessels. It may be possible to segment the data in real time to remove the lumen by detecting the start of the intima with amplitude-based thresholding. However, imaging flow in the lumen would also be beneficial for some applications, such as assessing patency of the primary vessel or imaging dissections and bifurcations [89] [36] [90].

Results indicated that SVD filtering provided superior separation of flow relative to the Butterworth filter in a mixed tissue environment containing both coherent and diffuse regions. This finding is consistent with reports that the SVD filter can separate bright targets based on slow time correlation that conventional frequency-based filters are not able to suppress [43]. In the case of the ex vivo vessel, the flow in the microtube was proximal to the tissue, i.e. there is little attenuation between the transducer and flow relative to the phantom case. For clinical imaging, blood between and vessel wall between the transducer and the vasa vasorum will provide greater attenuation relative to the ex vivo experiment, with attenuation due to blood expected to be ~6 dB at 40 MHz over 6 mm given α = 10 dB/cm at 40 MHz [91].

The observed improvements in Doppler CNR due to compounding in the ex vivo case were approximately 38% for both Butterworth and SVD filtering. The decreased improvement in Doppler CNR relative to the improvement in the phantom experiments was likely the result of higher SNR without compounding due to the described lower attenuation. Clinically, the blood flow signals would be attenuated by both blood in the lumen and by the intima.

Conclusion

Development of intraplaque vasa vasorum, an indicator of plaque vulnerability, is difficult to visualize using IVUS due to the difficulty of detecting slow flow in small vessels using ultrasound. Applying angular compounding between adjacent angular acquisitions increases sensitivity to flow in small vessels with slow flow. In experiments with flow at 5.6 mm/s in 200 μm tubes in tissue-mimicking phantoms, the Doppler CNR increased from 3.3 ± 1.0 to 13 ± 2.6 for Butterworth filtering and from 1.9 ± 0.15 to 7.5 ± 1.1 for SVD filtering due to angular compounding. When imaging flow at 5.6 mm/s in 200 μm tubes adjacent to the lumen of ex vivo porcine arteries, the Doppler CNR increased from 5.3 ± 0.95 to 7.2 ± 1.3 for Butterworth and from 23 ± 3.3 to 32 ± 6.7 for SVD filtering. In imaging the ex vivo porcine artery, flow was visualized in a 200 μm microtube even though the maximum backscattered amplitude of the vessel wall was 15 dB greater than that of flow within the tube. Translating these benefits to a high-rotation rate system could improve the ability to stratify risk and guide clinical decisions, such as determining which lesions require intervention in patients with stable disease.

• Separating flow from tissue in side-viewing IVUS is challenging but could aid in plaque characterization.

• Compounding adjacent angular acquisitions may increase sensitivity to slow flow in small vessels.

• Implementation of a correlation-based angular compounding approach increases Doppler CNR for rotational IVUS imaging.

• Approaches with both conventional clutter filtering and singular value decomposition filtering (SVD) are examined.