In vivo Visualization of Eye Vasculature using Super-resolution Ultrasound Microvessel Imaging

Abstract

Objective:

The choroidal vessels, which supply oxygen and nutrient to the retina, may play a pivotal role in eye disease pathogenesis such as diabetic retinopathy and glaucoma. In addition, the retrobulbar circulation that feeds the choroid shows an important pathophysiologic role in myopia and degenerative myopia. Owing to the light-absorbing retinal pigment epithelium (RPE) and optically opaque sclera, choroidal and retrobulbar vasculature were difficult to be observed using clinically accepted optical coherence tomography angiography (OCT-A) technique. Here, we have developed super-resolution ultrasound microvessel imaging technique to visualize the deep ocular vasculature.

Methods:

An 18-MHz linear array transducer with compounding plane wave imaging technique and contrast agent – microbubble was implemented in this study. The centroid intensity of each microbubble was detected using image deconvolution algorithm with spatially variant point spread function, and then accumulated in successive frames in order to reconstruct microvasculature. The image deconvolution technique was first evaluated in a simulation study and experimental flow phantoms. The performance was then validated on normal rabbit eyes in vivo.

Results:

The image deconvolution based super-resolution ultrasound microvessel imaging technique shows good performance on either simulation study or flow phantoms. In vivo rabbit eye study indicated that the micron-level choroidal and retrobulbar vessels around the optic nerve head were successfully reconstructed in multiple 2D views and 3D volume imaging.

Conclusions:

Our results demonstrate the capability of using super-resolution ultrasound microvessel imaging technique to image the microvasculature of the posterior pole of the eye. This efficient approach can potentially lead to a routinely performed diagnostic procedure in the field of ophthalmology.

I. INTRODUCTION

Characterization of the vasculature of the posterior eye would be of great clinical value for diagnosis and treatment conditions. The choroidal (a vascular network to supply the posterior layers of the retina with nutrients) structure as well as vascular functional changes, when superimposed, can allow clinicians diagnose and treat many eye diseases, such as diabetic retinopathy [1], age-related macular degeneration [2] and glaucoma [3]. Degenerative myopia is another common cause of visual impairment with irreversible degenerative changes of the posterior pole. The functional changes in retrobulbar and choroidal blood flow are likely to play an important pathophysiologic role in degenerative myopia [4]. Choroidal and retrobulbar blood flow are also hypothesized to be early biomarkers of degenerative myopia complications such as myopic choroidal neovascularization [5]. Moreover, the interaction of retinal, choroidal and retrobulbar vascular systems may play an important role in keeping the optic nerve head (ONH) healthy [6]. Thus, non-invasive assessment of deep ocular microvessel architecture and its early changes (i.e., ranges from few microns to tens of microns under different disease conditions [3, 7]) has been shown to be an important clinical need.

Optical coherence tomography angiography (OCT-A) has been emerged as a non-invasive technique to access the microvasculature of the retina at a high resolution. Due to the light-absorbing retinal pigment epithelium (PRF), it is more difficult to image choroid than retina via OCT-A. An improved swept-source OCT, utilizing a longer wavelength, has extended its capability to image choroidal vessel [8, 9]. However, its shallow penetration depth still limit OCT-A to image deeper choroidal region, especially for the high refractive error and long axial length in all forms of myopia. In addition, OCTA cannot image regions directly below the optically opaque areas such as retina with retinal prosthetic electrode array implants [10], and the retrobulbar vasculature beyond the sclera. An imaging modality, capable of imaging the posterior pole of the eye with high resolution is still desired.

Ultrasound (US) Doppler imaging has been clinically used to measure blood flow for the retrobulbar vessels supplying the eye. In order to achieve a deep penetration depth to observe posterior vasculature of the eye, the frequency range (10 MHz–20MHz) of the ultrasound transducer is commonly adopted. However, in the conventional line-by-line scanning mode, the frame rate of the US imaging is up to 50 frames per second, resulting in a relative low sensitivity in Doppler imaging. To improve the sensitivity, the technology of compounding plane wave imaging [11, 12] has been recently proposed for the potential applications of imaging anatomy and blood flow of choroid with high sensitivity. More specifically, in 2016, Urs et al. first investigated the feasibility of using ultrafast plane wave technique to image ocular anatomy and blood flow [13]. Later in 2018, Urs et al. characterized the choroidal blood flow in human healthy subjects in a horizontal plane superior to the optic nerve head at the resolution of few hundred micron [14]. Despite the level of a hundred micron spatial resolution was achieved by either conventional line-by-line method or plane wave imaging method at these frequency range, the spatial resolution is inherently constrained by the fundamental limits of diffraction, which defines the trade-off between resolution and penetration [15]. Owing to the small size of the microvascular structure of the posterior eye, which can be as low as few μm [16], a new imaging technique which can simultaneously image deep ocular vasculature while maintaining microscopic scale resolution is desired.

Similar to ultrasound imaging, spatial resolution limitations are of great concern in optical fluorescent microscopy. Sub-diffraction imaging has been developed to overcome the same fundamental physical limitations with a revolutionary improvement of resolution through the use of different technical approaches. Two major implementations, stochastic optical reconstruction microscopy (STORM) and fluorescence photo-activated localization microscopy (FPALM), exploit the stochastic blinking of many spatially isolated fluorescent particles within a diffraction limited region [17, 18]. A super-resolved image is obtained by accumulating all the localized centers of each separable source over thousands of acquisitions. The key localization process is usually done by finding the center or fitting a 2-D Gaussian to the emission intensity distribution of the fluorescent molecule. However, these analysis methods typically rely on isolated images of individual fluorophores, limiting the density of fluorophores that can be localized per frame, resulting in a longer time to reconstruct an image. As indicated by Mukamel et al [19], despite image overlap between molecules partially degrades the localization precision, the localization at a higher density of fluorophores is suggested to increase the overall imaging speed. As a consequence, image deconvolution based analysis is a valuable tool for fast superresolution microscopy at high density [20].

The average diameter of ultrasound contrast agents such as gas filled microbubbles (ranges between 1–3 μm) is considerably smaller than the US resolution limit, therefore, these spatially isolated microbubbles could be used to realize super-resolution (SR) techniques which are analogous to the results achieved the revolutionary technology of optical localization microscopy. Most of current SR ultrasound microvessel imaging, through localizing centroid of many spatially isolated microbubble signals, enables imaging of microvessels at resolutions as small as ten micrometers (an order of magnitude smaller than the ultrasound diffraction limit) [21–23]. Errico et al., first successfully reconstructed super-resolved image of rat brain vasculature network using ultrafast plane wave sequence. However, this method still requires a long data acquisition time of 150 seconds, which making this technology susceptible to motion artifacts. To move this technique a step forward, most recently, some motion correction algorithms [24] were applied to demonstrate its potential capability in the presence of patient motion. In addition, a more clinically applicable approach called motion model ultrasound localization microscopy [25] has been demonstrated in the preclinical and clinical multiparametric tumor characterization.

The U.S Food and Drug Administration (FDA) 510k standards for ophthalmic exposure are more stringent than for any other clinical specialty (i.e., mechanical index in FDA 510k limit is 0.23 for ocular tissues) [26]. Thus, a shorter data acquisition time and a lower mechanical index are preferred for imaging the ocular tissues. To achieve fast super-resolution imaging, Zhang et al. [27] proposed an acoustic wave sparsely activated localization microscopy (AWSALM) method to activate/deactivate decafluorobutane nanodroplets via acoustic switching. Despite they achieved super-resolved frames with sub-second temporal resolution, the mechanical index used in either AWSALM or fast-AWSALM [28] is still higher than the standard ophthalmic exposure. In addition, Bar-Zion et al. [29] utilized statistical methods, that is high order statistics to signal densities, to achieve fast vascular ultrasound imaging. As a result, they enhanced the spatial resolution with the reduced data acquisition time from densely grouped microbubbles. Despite a high temporal resolution was obtained, they only showed a small improvement in spatial resolution. Later, Yu et al. utilized an image deconvolution based technique, which is similar to the deconvolution method in the optical field, to obtain SR ultrasonic imaging with a high temporal accuracy [30]. It should be noted that the point-spread-function (PSF) of the imaging system was affected by various imaging depth and ultrasound transmission mode. The unique PSF used in previous studies may lead to limited accuracy of the image deconvolution method.

In this study, ultrafast plane wave imaging based SR microvessel imaging is first proposed for vasculature mapping of posterior pole of the eye using image deconvolution method. To improve the deconvolution process, pre-calculated spatially-variant PSF was chosen based on the locations of the imaging subjects. The accuracy of image deconvolution based method was first validated on a simulation study and then tested on experimental flow phantoms. Finally, multi-view images of rabbit eyes in vivo were performed to demonstrate the capability of our method to provide the vasculature network of the posterior pole of the eye, especially for choroidal and retrobulbar vessels.

II. MATERIALS AND METHODS

A. System setup

Figure 1 shows the experimental setup for SR ultrasound microvessel imaging. A high frequency Verasonics Vantage system (Verasonics Inc, Kirkland, WA, USA) and a linear array transducer L22–14v (Verasonics Inc, Kirkland, WA, USA) with a center frequency at 18 MHz and standard elevation focus at 6 mm were used. In order to adjust the distance between the array transducer and imaging targets, the array transducer was mounted on a three-axis translation motorized linear stage system (SGSP33–200, OptoSigma Corporation, Santa Ana, CA, USA) with a minimum step size of 60 nm in all three axes. Ultrafast plane wave data acquisition offers rich spatiotemporal information that can be advantageous for microbubble detection and tracking [31–33]. As a balance of image quality and frame rate, five different transmission angles (−6° to 6° with a step size of 3°) were implemented and then coherently added to produce one higher-contrast ultrasound image at 1 kHz effective pulse repetition frequency (PRF) in this study. All Time Gain Compensation (TGC) options were kept the same for all acquisitions. Delay-and-sum (DAS) based software beamforming was applied to raw RF data and the beamformed in-phase quadrature (IQ) data were saved for post-processing using MATLAB 2017a software (The MathWorks, Natick, MA, USA). All the system parameters were set identically in simulation, flow phantom and in vivo rabbit studies.

Fig. 1.

The experimental setup for SR ultrasound microvessel imaging. A 18 MHz linear array with 128 elements was fixed on a motorized 3-axis linear stage for 2D/3D scanning of a rabbit eye. Microbubbles were injected via ear vein access.

The pressure output of the array transducer was measured using a needle hydrophone (HGL-0085, ONDA Co, Sunnyvale, CA, USA) under the plane wave mode. To minimize the ultrasound-induced disruption of microbubbles, the transmitted pulses were 1 cycle sinusoids at 18 MHz with a peak rarefactional pressure of 220 kPa (mechanical index − 0.1). The measured acoustic pressures are within the FDA 510k standards for ophthalmic exposure.

B. Simulation Design

In simulation, ‘ground truth’ location of the image target was pre-defined, which allows us to compare it with the location reconstructed by image deconvolution based approach. The simulation mode of the same Verasonics Vantage system was used in this study where the attenuation is defined as − 0.5 dB/cm/MHz.

The system PSF at various depths were first simulated using single-point targets. Then, we simulated two-point targets to verify the image deconvolution approach. Figure 4 shows two-point targets at a depth of 6 mm (equal to the elevation focus of the array transducer). They are either separate by 100 μm in the axial direction or lateral direction, but kept at the same location in the other direction. Figure 5 (e–h) shows one two-point target laterally separated by 140 μm but at the same axial depth of 18 mm (maximum imaging depth for posterior pole of the rabbit eye). The axial or lateral separation distance was adjusted to take into account the calculated spatial resolution at each depth location to meet the Rayleigh criteria.

Fig. 4.

The simulated two-point target with a separation distance of 100 μm in (a-d) lateral direction and (e-h) axial direction at a depth of 6 mm. For each case, the raw data, 2D interpolated data, the 30-iterative intermediate result and its corresponding 1D intensity profile are shown. The final image deconvolution results are marked as the red crosses. The term “PSF-6” here refers to the PSF of a single-point target imaged at a depth of 6 mm.

Fig. 5.

The simulated two-point target with (a-d) lateral separation distance of 100 μm at a depth of 6 mm, and (e-h) lateral separation distance of 140 μm at a depth of 18 mm. For each case, 2D interpolated data, the 30-iterative intermediate result and its corresponding 1D intensity profile are shown. The final image deconvolution results are marked as the red crosses. The terms “PSF-6” and “PSF-18” here refer to the PSF of a single-point target imaged at a depth of 6 mm and 18 mm, respectively.

C. Flow Phantom Experiment

Before performing the flow phantom study, we first experimentally characterize the spatial resolution and spatially variant PSF at different depths. Since acquiring the exact PSFs of all microbubbles are impractical under experimental conditions, the PSF of the wire target, commonly treated as the system response of an ideal point source, has the potential to substitute for the PSF of the microbubble in the deconvolution process. To achieve this goal, a wire phantom consisting of one piece 10 μm of tungsten wire (California fine wire company, Grover beach, CA, USA) was used. The wire direction (x-y plane) was first positioned normal to the ultrasound image plane (x-z plane) so that the cross-section of the wire was estimated to be significantly smaller than the diffraction limited resolution of the scanner and was used to mimic a point scatter. Then the transducer was positioned at the center of the field-of-view (FOV) and was moving axially (z direction) to adjust the distance between the array and the wire phantom (up to 20 mm). Sufficient dwell time was allowed to ensure that the wire was stable before data acquisition. At each depth, a total of 320 ensembles were acquired. The full width half maximum (FWHM) through the pixel that contained the centroid was calculated for every individual frame and the imaging resolution of the ultrasound system was measured as the averaged FWHM of all 320 frames.

A home-made flow phantom with a 300 μm inner diameter and 640 μm outer diameter laboratory tube (Dow Silicones Corporation, Midland, MI, USA) was connected to a syringe pump (Model NE-1000, New Era Pump Systems Inc., Farmingdale, NY, USA) that produced a constant flow. The flow speed is set to 11.8 mm/s, corresponding to the average choroidal flow velocity [14]. Lumason (Bracco Diagnostics Inc., Monroe Township, NJ, USA) microbubbles (the mean diameter ranges 1.5 – 2.5 μm), a United States FDA approved clinical ultrasound contrast agent, was used in this study. Two concentrations which approximately correspond to 1×10⁸ (standard) and 1×10⁵ (low) microspheres/mL were used in the flow phantom. We investigated two representative locations − 6 mm and 18 mm, which were matched to the selected depths used in simulation. In each acquisition, 960 ensembles were collected, which is corresponding to 0.96 second ultrafast data acquisition.

D. Rabbit Eye in vivo

The in vivo rabbit experiment was performed according to the University of Southern California Institutional Animal Care and Use Committee (IACUC) protocol. Dutch Belted Pigmented rabbit (~2 kg) was single housed and fed with normal diet. Prior to the imaging experiment, the rabbit was anesthetized with ketamine (35 mg/kg) and xylazine (5 mg/kg) via subcutaneous injection. Two drops of phenylephrine were applied topically to prevent cornea swelling and decrease discomfort. Redosing of anesthesia was achieved by 2.5% sevoflurane through a facial mask. The heart rate, respiratory, body temperature, oxygen saturation, non-invasive blood pressure were recorded every 5 minutes until it became fully conscious.

Before applying ultrasound gel to couple the transducer to the surface of the eye, we used the retina camera (RetCam 3, Clarity Medical Systems Inc., Pleasanton, CA, USA) to visualize the posterior pole of the rabbit eye and roughly identify the orientation of the optic nerve head (ONH) as shown in Figure 2. The retinal vasculature confines itself to a pair of wing-shape areas that extend from the ONH both medially and laterally, which is consisted with previous study [34]. In this study, the orientation of the ultrasound transducer was placed at two different directions, one is parallel with the wing-shape areas while another is perpendicular to that region. Both of the scan planes crossed the retina/choroid somewhat obliquely in a horizontal plane superior to the optic nerve.

Fig. 2.

Before performing ultrasound imaging, we applied the retina camera to visualize the posterior pole of the rabbit, and roughly identify the orientation of the optic nerve head. A total of two cross-sectional scans were performed in this study. Note: the dash lines don’t represent the exact scan position of the US transducer, only indicate the scan direction.

During the US scanning, the real-time B-mode imaging and Doppler imaging (FlashDoppler.m in Verasonics system) were first implemented to find the region of interest (ROI). Then, a bolus of 0.3 mL standard concentration microbubbles was intravenously administrated through an ear vein via a 27 G syringe. In order to achieve a 3D volume imaging, the linear array transducer was initially tilted to the position parallel with US scan direction 2 in Fig.2 and then mechanically moved along US scan direction 1 by the motorized linear stage at the step of 60 μm. The step size is chosen based on the balance between the system resolution and the scanning time/plane numbers. A total of 100 scan planes were collected to reconstruct the 3D volume image.

E. Post-processing algorithm

For each experimental collected data, the beamformed IQ data was used to first obtain a raw B-mode image of the wire target. To reduce the computational complexity of interpolation procedure, we upsampled the wire data with a pixel size of 12.5 μm × 12.5 μm fine grid through a 2D spline interpolation function. To obtain the PSFs, intensities that was −20 dB or lower than the maximum intensity in 2D interpolation data were rejected in order to maintain the Gaussian shape. The 1D axial or lateral profile crossing the centroid peak intensity of the upsampled data was extracted and the spatial resolution was defined as the FWHM of each profile. Table 1 lists the calculated axial/lateral resolution of the conventional B-mode imaging and corresponding PSF kernels. According to the locations of flow phantoms, kernels (PSF-6 and PSF-18) at a depth of 6 mm and 18 mm were used in phantom study. In order to cover a large axial range of the posterior eye from choroid to retrobulbar vessels, two spatially-variant kernels (PSF-14 and PSF-16) at a depth of 14 mm and 16 mm were used.

Table 1.

The FWHM spatial resolution of the conventional B-mode imaging under experimental conditions and the corresponding selected PSF kernel for flow phantoms and posterior eye.

	Kernel	Imaging depth (mm)	Axial resolution (μm)	Lateral resolution (μm)
Flow phantom-1	PSF-6	6	98.03	121.14
		7	101.27	117.9
		8	100.85	119.13
Flow phantom-2	PSF-18	18	117.32	201.82
		19	119.29	212.49
		20	123.42	225.31
Posterior Eye	PSF-14	14–16	110	164
	PSF-16	16–18	114	183

The theoretical resolution limit of the super-resolution imaging technique can be calculated following the theories proposed in [23, 35]. More specifically, the axial resolution limit is given by equation (1) while the lateral resolution limit is determined by equation (2):

$${\sigma }_z\approx c{\sigma }_{\tau }/\left(2n^{1/2}\right)$$ (1)

$${\sigma }_x\approx 2\times 3^{1/2}c{\sigma }_{\tau }f/\left(D{n}^{1/2}\right)$$ (2)

Where c is the ultrasound speed, n is the number of channels used in receive processing, f is the focal length and D is the length of the transducer array. σ_τ is the timing resolution of the ultrasound system, which is bounded by the Cramer-Rao lower bound (CRLB) using the following equation [36]:

$$\begin{gathered} {\sigma }_{\tau }=\sigma \left(\widehat{{\tau }_0}-{\tau }_0\right) \\ \ge \sqrt{\frac{3}{2f_0^3{\pi }^{2}T\left(B^3+12B\right)}\left(\frac{1}{{\rho }^2}{\left(1+\frac{1}{SN{R}^2}\right)}^2-1\right)} \end{gathered}$$

where f₀ is the transmit pulse center frequency, T is the kernel size of for the time delay estimation, B is the pulse bandwidth, ρ is the normalized correlation coefficient between the experimental signal and the reference PSF used for localization, SNR is the signal-to-noise ratio of receive signals. Following these theoretical models, it is predicted that the 18 MHz array used in this study could attain an axial resolution limit of 1.7 μm, and lateral resolution limit of 5.5 μm at 6-mm depth, and 16.6 μm at 18-mm depth.

The details of image deconvolution based processing is illustrated in an example data set, shown in Fig. 3. A spatiotemporal singular-value-decomposition (SVD) [37] filter was first applied to remove background tissue and stationary microbubble signals. More specifically, all frames are used as a single window for spatiotemporal filter. The acquired IQ data was rearranged into a 2D Casorati matrix where one dimension is time and the other dimension is space. SVD filter could be performed on Casorati matrix, and give the corresponding temporal singular vector, spatial singular vector, more importantly – diagonal matrix. The diagonal matrix sorts the singular values which illustrate its relative contribution to the overall signal in a descendent magnitude. To extract the potential flowing microbubble signals, two threshold orders were selected [38]. The low-order threshold (i.e., where the slope degree becomes less than 45°) was used to reject the tissue signal, while the high-order threshold determined by the flattened curve slope was implemented to suppress the noise signal.

Fig. 3.

Processes of the image deconvolution method. (a) Example of a local region of SVD filtered microbubble raw data, followed by (b) 2D spatial interpolation. (c) threshold by intensity value < −20 dB of the maximum microbubble intensity value in each frame to further suppress noise. (d) 2D normalized cross-correlation maps with the selected PSF kernel. (e) intermediate result after 30 times iterations. (f) the final deconvolution output.

Then the filtered microbubble signal was thresholded by intensity value (i.e., rejecting pixels with intensity values less than −20 dB) to further suppress noise, followed by the same interpolation schemes mentioned above. Next, the spatially-variant PSF was used to perform a 2-D normalized cross-correlation with the interpolated microbubble signals in order to roughly identify the potential bubble regions. At each identified region, we applied a Richardson-Lucy algorithm which is a non-linear iterative deconvolution method [39, 40] on the original interpolated image for deconvolution process. The iterative deconvolution process is describe by equation (3) below

$$W_{i,r+1}=W_{i,r}\frac{1}{s_i}\sum _k\frac{s_{i,k}{H}_k}{\sum _j{s}_{j,k}{W}_{j,r}}$$ (3)

where W_r stands for the r^th reconstructed image, H_k represents the k^th pixel of the degraded images, S is the spatially-variant system PSF, r is the iteration step number, and s_i is the normalization factor for W_i. The final SR imaging of vasculature network was constructed by accumulating deconvoluted microbubble signals in successive frames, followed by the 3×3 median filter. For plotting 1D profiles of the line markers in rabbit eye, spline interpolation was further used to increase grid size to 2.5 μm.

III. RESULTS

A. Simulation results

Figure 4 shows SR imaging results of two-point targets with a designed distance of 100 μm in both the lateral (Fig. 4 a-c) and axial (Fig.4 e-g) directions. In either Fig. 4 (b) or Fig. 4 (f), the two-point target appears as a single-dot in the raw data, however, the characteristic shape and phase interference of it makes the separation of the two-point target possible. After applying the image deconvolution algorithm, we clearly observe that the two-point target was successfully isolated and identified as shown in Fig. 4 (c,g). The distances between the resolved two single points (red crosses) were calculated to be 99 μm for the lateral case and 97 μm for the axial case, which is consistent with the pre-defined value in simulation. Fig. 4 (d,h) shows the 1D profiles in both the axial and lateral directions. The presence of two local peaks in the intermediate deconvolution results, as opposed to only a single peak in the raw data, shows improved estimation of the true point locations, which demonstrated the efficacy of the image deconvolution algorithm.

Besides image deconvolution method itself, a correctly applied system PSF is critical as well. Fig. 5 displays the performance of the deconvolution process using two different PSFs, which are PSF-6 kernel and PSF-18 kernel. Fig. 5 (a–c) shows results for a simulated 100 μm laterally separated two-point target at a depth of 6 mm while Fig. 5 (e-g) shows results for a 140 μm laterally separated two-point target at a depth of 18 mm. After finding the deconvoluted centroid positions (red crosses), we measured the distance between each pair of red crosses. For the 6 mm target, we found a separation distance of 99 μm for the PSF-6 kernel while the PSF-18 kernel was not able to identify the difference. For the 18 mm target, the distance estimation was 172 μm for the PSF-6 kernel and 136 μm for the PSF-18 kernel. These results indicate that the correctly applied PSF kernel may be able to identify the centroid accurately. The mismatched kernel leads to estimation bias or missed points in the image, and this underscores the importance of proper kernel selection.

B. Flow phantom results

Low microbubble concentration can ensure a high probability of spatially isolated signals. Figure 6 displays a small region where only single microbubble is presented in the flow phantom. Fig. 6(a-c) presents the results at 6 mm depth without deconvolution, with kernel PSF-6 and PSF-18, respectively. At this depth, both kernels can correctly identify the bubble location. However, the performance was quite different for that at 18 mm depth. In Fig. 6(e), it can be clearly seen that after applying the PSF-6 kernel, the single bubble region was divided into three distinct ROIs. The two bubbles at the margin are artifacts caused by PSF-6. In Fig. 6(f), PSF-18 correctly identifies the single bubble with a shrinking dot instead of a large dot in Fig. 6(d). These observations are consistent with our simulation results. Figure 7 shows the full view of the flow phantom under low microbubble concentration. Within a short acquisition time, the structure of the flow phantom is not fully reconstructed in the SR image. In-plane flow velocity within the microvascular network was estimated by tracking microbubble positions over time via a Nearest-Neighbor method [41]. To reduce bias, the microbubbles that were tracked over at least 40 successive frames were analyzed and the velocity was calculated as the average distance divided by the time interval. The calculated velocity of 11.5 ± 0.6 mm/s, which is close to the designed value, demonstrates the accuracy of the image deconvolution approach.

Fig. 6.

Image deconvolution results of a small region (single microbubble) in the flow phantom under low microbubble concentration. Two different PSFs were used. (a–c) at the depth of 6mm. (d–f) at the depth of 18 mm. The final deconvolution results are marked as the red crosses.

Fig. 7.

The full flow phantom imaging results under low microbubble concentration. (a,b) at a depth of 6 mm and (c,d) at a depth of 18 mm. The first column represents the one frame of the SVD filtered raw data without interpolation. The second column is the accumulation of all deconvoluted points.

Figure 8 shows original B-mode image of the flow phantom and its corresponding SR microvessel imaging under standard microbubble concentration at the depth of 6 mm and 18 mm, respectively. It was observed that the flow vessels were fully reconstructed within a short data acquisition time (less than 1 second). Supplementary videos 1 and 2 show the accumulation process of the deconvoluted microbubble at the depth of 6 mm and 18 mm, respectively. Based on the reconstructed SR microvessel imaging, the average diameters of the flow phantom were found to be 294 ± 8 μm for 6 mm depth phantom and 290 ± 12 μm for 18 mm depth phantom, which fall within the range of precision of the inner diameter of the tube. However, the deep position phantom appears to have a slight higher variance.

Fig. 8.

The flow phantom imaging results at standard microbubble concentration. (a–c) phantom at a depth of 6 mm and (d–f) phantom at a depth of 18 mm. The first column represents the one frame of the SVD filtered raw data without interpolation. The second column is the accumulated microbubble localization map acquired less than 1 second in the study. The final column shows a zoomed-in view (white dash circle in 2^nd column) to provide a better understanding of the localized bubble distribution. Supplementary video 1 & 2 describe the full reconstruction process.

C. In vivo rabbit eye

Figure 9 shows the appearance of the post-processed scan of the rabbit eye in vivo. Fig. 9 (a-c) is the scanning direction across the wing-shape area centered on the ONH. The retina/choroid and sclera underneath are visualized in B-mode image. After filtering the stationary tissues and background noise, ultrafast plane wave based power Doppler image offers vasculature information with a high sensitivity - more vessel branches are revealed. It was observed that the quantity and density of vessels increase along the depth from the retina to choroid. The ONH region has a dense vasculature. However, its resolution is still insufficient to reveal the fine-scale retinal and choroid vessel, and the accurate vessel locations are difficult to identify in this mode. In the reconstructed SR imaging, the choroidal vessels, as the branches of the ciliary artery, are visible in fine scale. The sub-branch vessels in the choroid are indicated in the zoom-in view of SR imaging in Fig. 9(c). Regarding of the ONH region, it is convergence region of the ophthalmic artery, including short posterior ciliary artery and long anterior ciliary artery. Thus, the vessel distribution in conjunction with ONH have a relatively larger diameter, which are consisted with the eye structure. Fig. 9(d-f) shows another scanning direction of the posterior pole of the eye. In this view, the retrobulbar vessels beyond the sclera region are clearly observed. The retrobulbar vessels are distributed along the orbit, appearing as the circular shape. In the ONH region, the retrobulbar vessels go deeper following the direction of the optic nerve sheath. The cross-section profiles of the reconstructed SR vessels at three positions (line markers 1, 2, and 3) were shown in Fig. 10.

Fig. 9.

(a,d) The conventional B-mode images, (b,e) power Doppler imaging with microbubbles and (c,f) the reconstructed SR microvessel imaging of the posterior pole of the rabbit eye. (a–c) were collected at US scan direction 1 shown in Fig. 2 while (d–f) were collected at US scan direction 2 shown in Fig. 2. Three line markers were used to calculate the detectable resolution and the resolved distance. The white dash line points out a zoom-in view of the choroidal vessel.

Fig. 10.

Interpolated profiles across three line markers in the above Figure 9 (c,f).

The vasculature of the posterior pole of the rabbit eye was evidently visualized in multiple 2D SR microvessel imaging. Moreover, the 3D vasculature structure of the rabbit eye was obtained by reconstructing the 2D images from different scanning planes. Figure 11 shows the 3D volume imaging of the posterior pole of the eye. Fig. 11(a,b) indicated the power Doppler image and the SR microvessel imaging. Fig. 11(c,d) describes the corresponding vasculature image superposed at the original B-mode imaging. The vasculature network is consisted with its distribution in B-mode structural image. As indicated by the yellow arrow, the ONH region is a circular shape and some supporting vessels such as ciliary artery and central retina artery will go deep through the optic disk to the back side of the eye.

Fig. 11.

The 3D volume imaging of the posterior pole of the rabbit eye close to optic nerve head. (a) power Doppler imaging with microbubbles and (b) SR microvessel imaging. (c,d) are the corresponding vessel distribution superposed on the B-mode imaging.

IV. DISCUSSION

In vivo imaging of vasculature of the eye is a rapidly developing field in ophthalmology. Optical based imaging modalities such as OCT-A and Laser speckle flowmetry [42], are fundamentally limited by its shallow penetration depth and is susceptible to opaque tissues such as sclera and non-transparent cornea/lens. As a consequence, to image the deep choroidal vessel and retrobulbar vasculature beyond the sclera with a high resolution is still difficult. Ultrasound imaging has become an indispensable technique for ophthalmic imaging owing to its natural advantage of visualizing deep ocular structure through the use of high frequency ultrasound. Thus, the eye is potentially a fairly straightforward target for super-resolution ultrasonic microvessel imaging technology and could lead to quicker/easier clinical uptake. To the best of our knowledge, this paper presents the first report of using ultrafast plane wave imaging technique, with the contrast microbubble, which enables the observation of the deep choroid and retrobulbar vasculature network with a fine resolution.

An important step toward super resolution ultrasound microvessel imaging is to find the centroid of microbubbles in a finer grid. To achieve this goal, two major processing approaches have been previously proposed, including curve-fitting localization and image deconvolution approaches. More specifically, the curve fitting approach utilizes either 2D or 3D Gaussian profile to localize the centroids of each spatially isolated microbubble at a time, and the final SR image is then formed by accumulating all these localized points. Different from estimating the centroid locations, the image deconvolution method, processes the resolution-limited images iteratively and estimate the intensity in the final SR image. Foroozan et al. [43] has compared the performance of these two approaches in terms of localization accuracy and computational costs. According to their investigation, image deconvolution method has shown a much better accuracy than curve-fitting method. In addition, curve-fitting based localization method requires the spatially isolated microbubbles (i.e., a low microbubble concentration) with long acquisition times, which impedes its potential clinical applications because of the low temporal resolution. Thus, in this study, we adopted the image deconvolution algorithm to reconstruct the SR microvessel imaging under a very short data acquisition time. Both our simulation and experimental studies demonstrate the effectiveness of image deconvolution algorithm, which can shrink the bubble cloud and separate the bubbles. It should be noted that under standard microbubble concentrations, the image deconvolution based method has the increased possibility that fails to isolate multiple bubbles too close together, resulting in deconvolution bias or errors.

Previous image deconvolution based approaches only use one unique PSF of the ultrasound imaging system. However, due to the lack of transmit focusing, plane wave imaging is very susceptible to noise and its spatial resolution varies significantly. As shown in Table 1, the axial resolution ranges from 100 μm to 120 μm, while the variations in lateral resolution are unacceptably large. The worse lateral resolution at 18 mm is almost two times greater than that at 6 mm. Therefore, a fixed model based PSF is non-ideal for SR imaging localization. In this study, we applied spatially-variant PSF kernels which are pre-estimated by the various depths of the ROI region under the same transducer parameters. The simulation results in Fig. 5 and experimental flow phantoms using low microbubble concentration in Fig. 6 demonstrate that the spatially-variant PSF generates less artifacts and bias. More specifically, an inappropriately selected larger PSF can erroneously eliminate potential bubble signals while an incorrect smaller PSF can generate false-positive bubble artifacts. Under the standard microbubble concentration, the SR image of the flow phantom was reconstructed with a high temporal resolution. In addition, the reconstructed flow vessel is well consisted with our designed parameters such as the locations and vessel diameters. A slight higher estimation variance was observed at the deep flow phantom because of the worse resolution at deeper depth.

To demonstrate the feasibility of mapping the eye vasculature using the SR ultrasound microvessel imaging, we imaged the posterior segment of the rabbit eye in vivo. Due to the long axial distance of the posterior eye, two PSF kernels were used to process different regions of the rabbit eye. As expected, the SR ultrasound microvessel imaging provides a finer imaging resolution than power Doppler image. It has been documented that the translucent retina has less blood vessels and the choroid of the eye is primarily a vascular structure supplying the outer retina [44, 45]. And the oxygen and nutrition of the photoreceptor were provided by the choroid. Specifically, the choroidal vasculature is dense network with more vessel branches than retina. Our SR ultrasound microvessel imaging results confirmed that the deeper region – choroid has more vessels branches. In addition, both retinal and choroidal vessels are branches from the ciliary artery, and passing in conjunction with the ONH, resulting in a relative large vessel diameter. Moreover, the fine resolution of the retrobulbar vessels (the blind region for OCT-A) are first observed, which may provide additional diagnostic information for ocular diseases associated with microvasculature changes, such as the effect of intraocular pressure in glaucoma and degenerative myopia. Such a resolution may still not enough to reveal some retinal vessels at the size of few micron, and insufficient to distinguish between retina and choroid. However, SR ultrasound microvessel imaging offers a comparable spatial resolution but a deeper penetration depth with OCT-A. The spatial resolution of SR imaging can be further improved by the implementation of the linear array transducer with the optimum elevation focus to match the depth of posterior pole of the eye. All these results demonstrate that our SR imaging has the ability to provide the details of vasculature network of the posterior segment of the eye in fine scale resolution. This approach provides a potential way for practicing the ocular vasculature for human applications, especially for the region beyond the sclera.

There are few limitations to this study. One limitation is that image deconvolution based approach assumes a model with known PSF, which is typically approximated by a Gaussian curve. Despite spatially variant PSF has been considered in this study, the PSF is also influenced by the size of the microbubbles from two aspects. First, the microbubble size is not negligible compared to the fine grid dimension in SR imaging application. The PSF of the unique wire target is different from that of microbubbles with various radii and shape. These variance is difficult to control and not feasible to know all microbubble PSFs beforehand. Second, the microbubble radii (typically from 1 to 3 μm) affects the contrast to tissue ratio (CTR) of localization process. This is because that the resonance frequency of the microbubbles are located inside the conventional ultrasound transducer frequency from 1 to 10 MHz. However, a high frequency 18 MHz transducer is used in this study where most MBs are out of resonance, resulting in less effect on PSFs. In addition, the PSF used in this study is pre-calculated on the stationary wire target. In realistic situation, the microbubble is flowing with the blood in different directions. Brown et al. [46] investigated how CTR varies with microbubble flow speed and direction. However, owing to the lack of ground truth of the microbubble speed and direction in experimental condition, it will be challenging to correct for the introduced localization bias.

The second limitation is the detection techniques for super-resolution of microvasculature. There are several methods for microbubble detection, including pulse inversion [47], differential imaging [48] and SVD filter. Considering the high frequency was used in this study, we preferred SVD filter by taking spatiotemporal information into consideration. However, the localization accuracy of SVD filter is still affected by the flow speed. Brown et al. [46] demonstrated that SVD filter is unsuitable for stationary microbubble and may generate a localization error on the order of hundreds of microns when the flow velocity is less than 2 mm/s. In addition, the stack size and overlap has been shown to affect contrast-to-tissue ratio of the filtered images. Although there is still no universally agreed stack size (much of the recent literature does not clearly state the size being processed), a larger block size would gradually suppress the noise level while could blur the blood signals as well. An optimal stack size with a balance between the noise suppression and useful blood signal will be investigated in the future study to improve the visualization of the vasculature of the posterior pole of the eye.

Another limitation is the experimental validation of the achievable resolution of the proposed method. Owing to increasing chances of signal interferes (multiple scattering) under standard concentration, simple deconvolution method and shorter acquisition could potentially reduce the expected spatial resolution. Despite a single vessel phantom has been implemented to validate the proposed method in this study, the deconvolution bias from multiple bubbles (too close to each other) has less effect on the reconstruction results because their centroids are still in the vessel. In the future study, we would experimentally separate two structures with a known distance below the diffraction limit under standard concentration setup to accurately validate the achievable resolution.

V. CONCLUSIONS

In summary, our study first demonstrated SR ultrasound microvessel imaging of the posterior pole of the eye at high resolution, especially for deep choroidal and retrobulbar vasculature network. In addition, we implemented an image deconvolution method to take advantage of spatial information of various PSF, with the objective of improving the overall robustness of SR microvessel imaging. Owing to the limited imaging time available in clinical settings, this efficient method shows promise for future translational study and development, especially on deep ocular microvasculature including choroidal and retrobulbar vessels.