A Wearable Dual-Mode Probe for Image-Guided Closed-Loop Ultrasound Neuromodulation

Abstract

Low-intensity focused ultrasound (FUS) is an emerging non-invasive and spatially/temporally precise method for modulating the firing rates and patterns of peripheral nerves. This paper describes an image-guided platform for chronic and patient-specific FUS neuromodulation. The system uses custom wearable probes containing separate ultrasound imaging and modulation transducer arrays realized using piezoelectric transducers assembled on a flexible printed circuit board (PCB). Dual-mode probes operating around 4 MHz (imaging) and 1.3 MHz (modulation) were fabricated and tested on tissue phantoms. The resulting B-mode images were analyzed using a template-matching algorithm to estimate the location of the target nerve and then direct the modulation beam toward the target. The ultrasound transmit voltage used to excite the modulation array was optimized in real-time by automatically regulating functional feedback signals (the average rates of emulated muscle twitches detected by an on-board motion sensor) through a proportional and integral (PI) controller, thus providing robustness to inter-subject variability and probe positioning errors. The proposed closed-loop neuromodulation paradigm was experimentally demonstrated in vitro using an active tissue phantom that integrates models of the posterior tibial nerve and nearby blood vessels together with embedded sensors and actuators.

I. Introduction

THE use of low-intensity focused ultrasound (FUS) to modulate the activity of peripheral nerves [1], [2], [3], [4] is becoming increasingly common due to its non-invasive nature, cost-effectiveness, and precise spatial/temporal localization [5], [6], [7]. FUS has both thermal and non-thermal effects on tissues; the former are due to localized tissue heating, while the latter are driven by tissue displacements that result in mechanical actuation of voltage-gated ion channels in the cell membrane and subsequent production of synaptic transmission and motor excitation in the nervous system [8], [9]. Such tissue displacements can be directly measured using high-frequency ultrasound imaging [9], thus allowing direct feedback control of the modulation parameters.

One example clinical application of peripheral neuromodulation is percutaneous tibial nerve stimulation (PTNS) [10], [11]. PTNS is a promising treatment for overactive bladder (OAB), commonly affecting senior people. PTNS currently requires a clinician to administer the invasive percutaneous electrical stimulation. Thus, replacing the corresponding electrodes with FUS from an external array is of great interest for future at-home treatments. FUS is also promising as a non-invasive replacement for electrical vagus nerve (VN) modulation [12], [13], which is used to treat epilepsy, heart failure, and psychiatric conditions while eliminating the risks of implanted modulation electrodes such as voice alterations and dyspnea [14], [15]. Other potential applications are treatment of neuropathic trunk and limb pain [16], which have limited therapeutic options today.

Such applications require chronic and patient-specific stimulation, preferably accessible in at-home settings, which suggests the use of wearable and closed-loop FUS devices [17], [18], [19]. Several research-based or commercialized fully integrated wearable ultrasound devices for health monitoring have been developed [20], [21], [22]. However, they are either not designed for ultrasound imaging [21], [22] or not well-suited for image-guided neuromodulation of relatively small targets, such as the tibial nerve, due to low signal-to-noise ratio (SNR) and long acquisition times for low-power synthetic aperture imaging using linear arrays [20]. Even though our image-guided neuromodulation approach does not involve a fully integrated wearable device, it is capable of acquiring real-time images with high temporal resolution and good contrast-to-noise ratio (CNR), which are critical for image guidance on a human subject to minimize safety risks. Fig. 1 illustrates our approach, in which neuromodulation is performed on a human subject using a wearable dual-mode ultrasound probe wrapped around the body part of interest and connected to a transceiver and controller. Functional signals are fed back to the controller from the modulation target to optimize modulation parameters. This paper focuses on the design of the wearable probe.

Fig. 1.

Overview of the proposed image-guided neuromodulation approach, with tibial nerve stimulation used as an example.

Research on wearable ultrasound probe design and fabrication has followed different paths. One path focuses on transforming conventional rigid transducer arrays into mechanically flexible and body-conformal arrays for diagnostic and therapeutic ultrasound. Traditional focused ultrasound transducers use water-filled “coupling cones” to obtain good acoustic impedance matching to tissue, but these are unsuitable for wearable devices. In contrast, flexible arrays can wrap around complex curved surfaces (e.g., core regions such as the abdomen, or extremities such as legs, ankles, or fingers) to obtain impedance matching via direct mechanical contact. The resulting partially-wrapped or fully-wrapped configurations are robust to body motion and also have improved angular coverage and comparable image quality [23], [24], [25]. However, despite their optimal conformability to human bodies, such devices still face several challenges before they can be employed in clinical practice. Flexible materials (e.g., PVDF or other soft polymers) have low electromechanical coupling efficiency [26], which limits their use in energy-constrained wearable devices. Flexible arrays are susceptible to breakage during long-time use (although this may be mitigated as devices reach greater stages of clinical maturity) [25], [27]. The focal points of flexible phased arrays are also indeterminate for neuromodulation on curved surfaces with an unknown radius of curvature. Prior work has proposed several methods to estimate the shape of flexible transducer arrays and then calculate accurate time delay values for ultrasound beamforming. However, these approaches are unsuitable for continuous and real-time ultrasound biomedical applications due to the need for i) external positioning or optical sensors to track the array element positions [28], [29], and ii) iterative processing of various optimization algorithms involving entropy, gradient and variance [30], [31], [32]. Linear arrays avoid these issues at the cost of lower resolution, which results in reduced sensitivity for small targets. Also, image formation using flexible arrays is prone to distortion and blurring due to varying array geometry on curved surfaces [33], [34]. As an alternative, wearable ultrasound probes have been fabricated by mounting rigid transducer arrays on flexible substrates or sticky acoustic coupling materials [35]. The latter need to provide strong adhesion between the array and skin, along with good acoustic impedance matching and low acoustic loss.

Other challenges in delivering therapy to specific targets outside of the brain (e.g. the spinal cord, peripheral nerves, or organs themselves) include i) variability in the experimental setup, and ii) motion of internal tissues and organs (intrafractional motion) [36]. Respiration is the main cause of intrafractional motion, but variable filling of the rectum and bladder, peristalsis, and cardiac motion are also significant. Simple methods for limiting such motion include i) patient immobilization; ii) shallow breathing; and iii) abdominal compression [37]. However, such methods are unsuitable for ambulatory patients and/or outside a controlled clinical environment.

Another challenge is to determine optimal values for the neuromodulation parameters (frequency, repetition rate, waveform, and power level). These are expected to be both patient-specific (due to biological heterogeneity) and time-dependent (due to tissue/nerve remodeling and homeostatic responses such as habituation). Similar challenges are well-known for electrical stimulation, for which the dominant “patient fitting” paradigm is based on human-in-the-loop optimization under the supervision of a clinician. However, this process is slow and often fails to find the true optimum. Thus, there is growing interest in automated optimization via closed-loop devices that combine sensing with stimulation, e.g., for deep brain stimulation (DBS) [38], [39], [40] or spinal cord stimulation [41], [42]. However, closed-loop FUS remains largely unexplored.

This paper proposes wearable dual-mode ultrasound probes that can be wrapped around appropriate body surfaces (e.g., the neck and ankle for vagus and tibial nerve modulation, respectively) for image-guided closed-loop FUS therapy. Even though our wearable probe is not as body-conformal as typical flexible transducers, it provides a workaround to bypass most of the aforementioned issues by leveraging the natural deformability of human skin. The proposed design overcomes the challenges of flexible transducer arrays along with the scaling issues of manually assembling individual transducer elements onto a flexible substrate, as in our previous work [27]. Instead, the probes are designed to be rigid but assembled on a flexible printed circuit board (PCB), which integrates two transducer arrays optimized for imaging and neuromodulation, respectively. The former array acquires B-mode ultrasound images to locate the target nerve, thus allowing the latter to be adaptively focused on the target to maximize the delivered pressure. Small rigid arrays are acceptable for wearable probes because deformation of the human skin still allows for good interfacing of the probe with curved body surfaces. Moreover, a functional feedback mechanism is developed and verified in vitro on an ankle phantom with electromechanical nerve and muscle models, using a motion sensor to measure the FUS-induced activation of the selected artificial skeletal muscles. Even though the closed-loop neuromodulation system is demonstrated in vitro setup, the setup is realistic in terms of the time-scale and algorithms applied to emulate the mechanisms of neuronal firing and subsequent motor excitation of the innervated muscle fibers during neural stimulation. The motion signature sensed from the ankle surface is also a valid feedback signal, as evidenced by observations of ankle or joint motions triggered by stimulation of the peroneal nerve [43] or the tibial nerve [44].

The rest of the paper is organized as follows. Section II describes the design and fabrication of the ultrasound arrays used by the proposed dual-mode FUS probe, while Section III describes the assembled probe and its experimental characterization. Use of the probe for target localization and closed-loop FUS is discussed in Section IV and Section V, respectively. Experimental results from a closed-loop FUS prototype are described in Section VI. Finally, Section VII summarizes our contributions and concludes the paper.

II. Wearable Ultrasound Probe Design

In this section, we first present the design of our custom ultrasound imaging and modulation transducer arrays made from a standardized piezoelectric composite material. We then illustrate an ultrasound probe assembly specifically designed to optimize the integration of the probe with the human body.

A. Transducer Array Design

Fig. 2 compares the mechanical design of our ultrasound probe with other body-worn designs. Fig. 2(a) shows a conventional design in which both the transducer array and the substrate are rigid, such that the probe is not body-conformal. Fig. 2(b) improves wearability by using a a flexible acoustic couplant with strong adhesion to connect a rigid probe to the body [35]. Fig. 2(c) shows a body-conformal probe in which both the array and substrate are flexible [23], [24], [25]. Finally, Fig. 2(d) shows our approach, which features dual-mode rigid transducer arrays bonded to a flexible PCB substrate that can be wrapped around the body surface. Compared to the other body-conformal approaches, our probe is easy to design and fabricate since it does not use either a low-loss flexible acoustic couplant or a flexible transducer array. Instead, the only flexible component is the substrate, which is designed and fabricated using standard low-cost flexible PCB technology.

Fig. 2.

Different types of body-worn ultrasound probe designs. (a) Rigid transducer array bonded to a rigid substrate. (b) Rigid transducer array bonded to a flexible acoustic couplant. (c) Flexible transducer array bonded to a flexible substrate. (d) Our approach, which uses a rigid transducer array bonded to a flexible substrate.

The operating frequencies of the imaging and modulation arrays in our design were separately optimized since they have different performance goals and constraints. For the imaging array, the choice of frequency involves a trade-off between high lateral resolution, large penetration depth, minimal grating lobe artifacts, minimum pad width/spacing constraints from PCB manufacturing, and the complexity of bonding the array to the PCB. A nominal frequency of 4 MHz was chosen to balance between these competing factors. For the modulation array, the choice of frequency involves a similar trade-off between lateral resolution (neuromodulation selectivity), penetration depth, neuromodulation effectiveness (which is a strongly increasing function of the mechanical index) and supply voltage restrictions of the ultrasound transmitter. A nominal frequency of 1.3 MHz was chosen to balance between these competing factors.

The design of the 4 MHz imaging and 1.3 MHz modulation transducer arrays is demonstrated in the lower panel of Fig. 3. The 4 MHz array is a 1-D phased array with 64 transducer elements. The array has a kerf of 0.1 mm and a pitch size of 0.25 mm, which satisfies the criterion of half-wavelength pitch to minimize grating lobes that can possibly cause image artifacts. For the 1.3 MHz array, the pitch size and kerf are set at 1.2 mm and 0.2 mm. The modulation array contains only 16 transducer elements, which is expected to provide high enough acoustic pressure to modulate the activity of peripheral nerves (such as the Tibial nerve) at shallow focal depths of ~20 mm using transmit focusing. The array design follows the same rule of half-wavelength pitch to produce ultrasound beams without grating lobes, such that acoustic pressure is maximized only at the focal point when FUS is implemented. In order to maximize acoustic pressure levels during stimulation, the transverse height of each transducer element in the modulation array is configured to be 10 mm, which is approximately double that of each imaging array element along the same dimension. Moreover, based on the operating center frequency of the two transducer arrays and the frequency constant of the PZT material in the thickness vibration displacement mode, the corresponding thicknesses of the 1.3 MHz and 4 MHz transducer arrays are set to 1.54 mm and 0.5 mm, respectively. To maximize the acoustic coupling efficiency of the dual-mode probes, we use rigid PCBs with adjustable thickness placed between the array assemblies and the flexible PCB substrates. This compensates for the difference in the thickness of the two array assemblies operating at different center frequencies, thus ensuring a flat surface when applied on the human skin. Also, based on the flexible PCB specifications provided by the manufacturer, the minimum bending radius of the flexible part of the probe is 10–15 times the PCB thickness, i.e., approximately 3 mm for our PCB design. This value is much smaller than the radius of curvature required to wrap around a typical human body part such as the neck or ankle. Therefore, the proposed wearable probe is suitable for repetitive bending and continuous use in multiple neuromodulation trials.

Fig. 3.

Design of the wearable ultrasound probes. The 1.3 MHz and 4 MHz rigid transducer arrays are assembled on a flexible PCB substrate for ultrasound imaging and modulation. The top panel of the figure shows cross-sectional views of the ultrasound probe assemblies, including the electrode wrap-around design of the transducer arrays. The bottom panel shows the bottom view of the two transducer arrays.

Both transducer arrays were fabricated using a piezoelectric composite material based on lead zirconate titanate (PZT-5A) and an epoxy filler and manufactured using a dice-and-fill method. A variety of composite geometries are possible, as shown in Fig. 4. Each geometry is characterized by a pair of numbers of the form $n_1-n_2$, where $n_1$ and $n_2$ are the number of continuous dimensions for the active (piezoelectric) and passive (filler) components of the composite, respectively. Here, we selected a 2–2 PZT composite material, as shown in Fig. 4(c), based on its lower cost, better electromechanical performance, and wider −6 dB bandwidth compared to 1–3 composites [45]. Note that the term “2–2” indicates that both active PZT and passive epoxy materials are self-connected along two dimensions, namely $x$ and $z$.

Fig. 4.

Possible geometries of piezoelectric composite materials: (a) 0–3, (b) 1–3, and (c) 2–2. Coordinate axes for imaging are indicated.

Exploiting a PZT composite material to fabricate transducer arrays can provide the benefits of lower acoustic impedance, better mechanical flexibility, and lower mechanical quality factor [46]. The acoustic impedance, $Z_A$, is reduced by a factor equal to the fraction of total volume occupied by the PZT material, which is known as the fill factor. Given the acoustic impedance of the selected bulk PZT material ($Z_A\approx 35$ MRayls), the acoustic impedance of our 4 MHz and 1.3 MHz transducer arrays can then be calculated to be approximately 21.1 MRayls and 29.5 MRayls, as summarized in Table I. The reduced acoustic impedance enables better impedance matching with the human skin ($Z_A\approx 1.48$ MRays), which in turn enhances both imaging sensitivity and the acoustic intensity available for FUS stimulation.

TABLE I.

Transducer Array Parameters

Transducer Property	Imaging	Modulation
Array type	Phased array	Phased array
Number of elements	64	16
Active aperture	19 mm	15.9 mm
Frequency	4 MHz	1.3 MHz
Array thickness	0.5 mm	1.54 mm
Array element width	0.15 mm	1 mm
Kerf	0.1 mm	0.2 mm
Array material	PZT	PZT
	2-2 composite	2-2 composite
Composite fill factors	0.6	0.84
Array $Z_A$	29.5 MRayls	21.1 MRayls
Matching layer thickness	0.19 mm	0.60 mm
Matching layer material	Epoxy, ${\text{SiO}}_2$	Epoxy, ${\text{SiO}}_2$
Matching layer $Z_A$	4.8 MRayls	4.8 MRayls
Rigid PCB thickness	1.6 mm	0.6 mm
Flexible PCB thickness	0.26 mm	0.26 mm

The upper panel of Fig. 3 depicts the patterns of the wrap-around copper electrode on the two transducer arrays. The signal electrodes of either array are patterned at the bottom of individual elements to electrically connect to the transceiver circuit. The ground electrode is designed in an inverted U-shape with small parts extended to the array bottom to electrically connect with ground pads on the circuit board. The effective area of piezoelectric transduction consists of the PZT areas that are sandwiched between the signal and ground electrodes.

B. Wearable Probe Design

The tops of the transducer arrays are covered with an impedance matching layer to minimize reflections at the transducer-skin interface and thus maximize sensitivity for both ultrasound imaging and neuromodulation. The chosen material for this layer is a mixture of epoxy and silicon dioxide (${\text{SiO}}_2$), which results in an acoustic impedance of $Z_A\approx 4.8$ MRayls that is close to the optimum value of $\sqrt{Z_{A1}{Z}_{A2}}$ for a quarter-wavelength impedance transformer. Accordingly, the thickness of each matching layer is configured to be a quarter wavelength at the nominal operating frequency.

Making the ultrasound probe wearable requires compensation for the difference in total thickness (transducers plus matching layer) of the two arrays, such that both of them can simultaneously interface with the body surface. For this purpose, both rigid transducer arrays are first assembled on small rigid PCBs. The thicknesses of these carrier PCBs are configured to be 1 mm apart, as shown in Table I, to ensure equal thicknesses for both array assemblies. The carrier PCBs are then assembled on a single flexible PCB with a 0.26 mm-thick polyamide substrate, thus realizing a wearable and body-conformal dual-mode probe. All the parameters of the imaging and modulation arrays (at 4 MHz and 1.3 MHz, respectively) used by the wearable probe are summarized in Table I.

III. Probe Characterization

The sensitivity, bandwidth, and acoustic pressure profiles of the assembled dual-mode ultrasound probes were measured in a water tank by using a highly sensitive (282 nV/Pa), waterproof capsule hydrophone (Onda HGL-0400) and a 20 dB gain pre-amplifier. During the measurements, the hydrophone was immersed in water and pointed towards the surface of the transducer arrays to be characterized. The performance of both modulation and imaging arrays was then characterized by focusing and steering the ultrasound beam. The necessary time delays of the array channels were configured using a research ultrasound platform (Verasonics Vantage 64LE) in pulsed mode. The transmit voltage was also configured over the available range $\left(0-100{\text{V}}_{pp}\right)$ through the Verasonics platform. The hydrophone was positioned 20 mm away from the arrays by default, and the pre-amplified voltage waveforms were recorded using a digital oscilloscope.

The spatial resolution and contrast-to-noise ratio (CNR) of the imaging array were evaluated separately. For this purpose, the array was used to generate B-mode images of a tissue-mimicking ultrasound calibration phantom (model 040GSE, CIRS) using coherent plane-wave compounding (CPWC). The choice of CPWC as an imaging method was driven by its ability to obtain comparable CNR to traditional focused imaging with a much smaller total number of scans, which is beneficial for saving energy and thus extending the operating lifetime of portable and wearable ultrasound systems.

A. Modulation Probe Characterization

The frequency response of the modulation array was measured at a transmit voltage of $40{\text{V}}_{pp}$ and a focal depth of 20 mm. The results are shown in Fig. 5(a) over a center frequency range of 0.5 MHz to 3 MHz. Note that the drive voltage was limited to $40{\text{V}}_{pp}$ by the maximum available transmit voltage of the Verasonics platform at frequencies below 1 MHz. The result shows a −3 dB bandwidth of ~700 kHz, corresponding to a quality factor of $Q\approx 1.86$.

Fig. 5.

(a) Frequency response measurement of the modulation array. (b) Sensitivity of the modulation array.

In addition, the sensitivity of the modulation array at the nominal center frequency (1.3 MHz) was characterized by measuring the peak acoustic pressure with the hydrophone at the same position but with the transmit voltage swept from $10{\text{V}}_{pp}$ to $100{\text{V}}_{pp}$. As shown in Fig. 5(b), the peak pressure increases linearly with the drive voltage, reaching a maximum of ~1.5 MPa. The peak pressure level from the modulation array (again operated at 1.3 MHz) was also measured over a range of focal depths from 20 mm to 50 mm. The result, which is plotted in Fig. 6(a), shows a gradual decrease in pressure with focal depth, in agreement with theoretical predictions.

Fig. 6.

Simulation and measurement results of the modulation array for a single focal point using FOCUS [47]: (a) measured and simulated normalized acoustic pressure and lateral focal spot width (full width half maximum, FWHM) as a function of focal depth; (b) normalized acoustic pressure and lateral focal spot width (FWHM) at the focal point as a function of beam steering angle.

Due to the unavailability of a motion control stage capable of precisely moving the probe with micron-level resolution, only simulations were employed using the FOCUS ultrasound simulator [47] to accurately estimate i) the ultrasound lateral focal spot width; and ii) the acoustic pressure for different values of focal depth and steering angle. For convenience, the simulated acoustic pressure amplitudes were normalized to those obtained for a focal depth of 20 mm and a steering angle of 0°. The simulation results, which are summarized in Fig. 6(a), demonstrate a decrease in the acoustic pressure and a rise of the lateral focal spot width, as computed through the full width at half maximum (FWHM), when the focal depth is increased from 20 mm to 50 mm. A similar trend is clearly seen in Fig. 6(b) when the beam steering angle is set from 0° to 40°. The relatively small lateral size of the focal spot (width < 2 mm for steering angles < 30°) allows for highly selective neural stimulation. For example, the resolution is good enough to selectively excite different groups of nerve fibers within the tibial nerve, which has a typical diameter of ~4 mm in adults [48].

B. Imaging Probe Characterization

Fig. 7(a) shows the frequency response of the imaging array with the operating frequency swept from 2 MHz to 6 MHz. The response shows a large −3 dB bandwidth of ~3 MHz, indicating an overdamped transient response and good axial resolution. The axial resolution of the imaging array was also directly estimated by characterizing the spatial pulse length (SPL) of the transmitted pressure waveforms.

Fig. 7.

(a) Frequency response measurement of the imaging array. (b) A one-cycle transmit pulse and the corresponding pulse measured from the imaging array when ultrasound imaging is implemented using a single plane-wave transmit pulse. (c) Zoom-in view of the transmit pulse waveform. (d) Zoom-in view of the measured pulse waveform.

Fig. 7(b)–7(d) show i) the waveform of a one-cycle bipolar ultrasound pulse transmitted through the imaging array at a voltage of $100{\text{V}}_{pp}$ from the imaging array during plane-wave imaging; and ii) the corresponding pressure waveform after conversion to voltage using a hydrophone. Two half-width equalization pulses were added to the beginning and end of the main single-cycle pulse to minimize the DC component of the transmitted spectrum. The ringing that is visible immediately after the ultrasound pulse is probably caused by i) the lack of backing layers, and ii) acoustic impedance mismatch at the transducer-matching layer and matching layer-tissue interfaces. The reflected waves arising from these mismatches cancel each other when the matching layer has a thickness of $\lambda /4$, i.e., at the nominal center frequency of 4 MHz. Thus, the reflections correspond to off-resonance components of the transmit spectrum.

From Fig. 7(b), the temporal pulse width including ring down time was estimated to be $t_p\approx 2.06\mu \text{s}$. Assuming water as the medium where the ultrasound waves propagate during the measurement, the axial resolution can then be calculated as $\text{SPL}/2=c{t}_p/2\approx 1.55\text{mm}$, where $c\approx 1500\text{m}/\text{s}$ is the speed of sound.

To ensure an accurate measurement, a small hyperechoic point target within the calibration phantom (diameter = 2 mm) was used as an imaging target to experimentally determine both the axial and lateral resolution of the imaging array, whereas a larger hyperechoic circular target (diameter = 10 mm) was employed for estimating the CNR. Fig. 8(a) depicts how the operating frequency affects spatial resolution and CNR within the array bandwidth of around 3 MHz to 5.5 MHz. The mean axial resolution (including error bars) monotonically decreases from ~1.83 mm to ~1.41 mm due to decreasing wavelength and SPL, assuming that the number of pulse cycles is fixed. The mean lateral resolution decreases from ~1.44 mm to ~1.14 mm, as expected for a largely diffraction-limited system. In addition, Fig. 8(b) shows that the mean CNR stays at a high level of 26–29 dB within the bandwidth of the imaging array, but drops rapidly outside this range due to reduced sensitivity and SNR. Finally, Fig. 8(b) also shows that the mean CNR gradually increases as the number of steering angles increases from 0° to 32°. This result quantifies the ability of CPWC to average out speckle, which in turn improves image contrast in scattering media.

Fig. 8.

B-mode imaging results from the imaging array using CPWC: (a) axial and lateral resolution using 32 steering angles as a function of the center frequency; (b) contrast-to-noise ratio (CNR) as a function of the center frequency and the number of steering angles. Each measurement was repeated 10 times (for (a)) or 5 times (for (b)) to ensure consistent results.

IV. Target Localization Using B-Mode Imaging

In this section, we first use the same CIRS calibration phantom to study the B-mode imaging performance of the custom wearable imaging probe. Next, we use the probe to image body parts relevant for peripheral neuromodulation (such as the human neck and ankle). All the human subject data were obtained from members of the the research team who provided informed consent. For this purpose, the arrays were connected to the same Verasonics platform. The CPWC plane-wave imaging technique with a sector scan was leveraged to achieve ultrasound imaging with good image quality (i.e., resolution and CNR) and a high frame rate.

Reference images of the calibration phantom and human tissues were obtained at 4 MHz using a commercial rigid probe (Verasonics L11–5v). The latter is a 128-element linear array with an element pitch of 0.3 mm and a −3 dB bandwidth of approximately 4.5–10.5 MHz. However, only the central half (consisting of 64 elements) was used during the experiments to facilitate comparison with the custom 64-element imaging array.

A. B-Mode Imaging Results

Fig. 9(a) and 9(b) demonstrate typical B-mode images of several circular hyperechoic targets of different sizes within the calibration phantom, as generated by both the custom imaging probe and the commercial probe (L11–5v). Firstly, note that the image acquired by the custom imaging probe does not display all the targets visible in the image acquired by the L11–5v probe in Fig. 9(b) due to its smaller aperture and field of view. Nevertheless, the objects visible in Fig. 9(a) agree with the expected target dimensions and positions of the phantom. In addition, the image resolution and CNR are comparable to those measured from the image acquired by the L11–5v probe when operated at a center frequency of 4 MHz. Quantitatively, the lateral resolution of the wearable probe is close to that of the L11–5v probe but the axial resolution is approximately 2× worse, as illustrated in Table II. The difference in axial resolution is likely due to the lack of backing layers in the custom probe, which results in a longer ring down time and increases SPL. Additional factors may include i) the use of a 2–2 PZT composite instead of a 1–3 composite, which typically contains more damping material (epoxy) to reduce ring down time; and ii) imperfect impedance matching between the composite and matching layer at 4 MHz. However, the CNR values of the two images are approximately equal when the same number of CPWC steering angles (32 in this case) are used. The imaging performance of the wearable probe on the hypoechoic dark targets is also quite close to that of the L11–5v probe, as visible in Fig. 9(c) and 9(d).

Fig. 9.

B-mode images of a calibration phantom obtained using the wearable imaging probe and the L11–5v probe. CPWC with 32 uniformly-distributed steering angles was used in both cases. (a) and (b) Bright (hyperechoic) targets. (c) and (d) Dark (hypoechoic) targets.

TABLE II.

Imaging Probe Resolution and Contrast-to-Noise Ratio

Probe Type	Frequency (MHz)	Axial Resolution (mm)	Lateral Resolution (mm)	Contrast-to-Noise Ratio (dB)
Custom imaging array	4	1.56	1.46	28.37
Commercial L11-5v linear array	4	0.67	1.02	32.19

Fig. 10 shows typical B-mode images acquired from the vagus nerve region on the left side of a human subject's neck using the wearable and L11–5v probes separately. Both the common carotid artery (CCA) and internal jugular vein (IJV) can be clearly detected and localized from the two images at similar positions. These imaging results suggest that the wearable probe can successfully detect the positions and orientations of major blood vessels. This information can then be used as anatomical landmarks to guide the process of localizing nerves (such as the vagus and tibial nerves). This indirect approach is generally much easier to implement than trying to directly detect nerves from the relatively low-resolution B-mode images generated by the wearable probe.

Fig. 10.

B-mode images of the left side of a human neck obtained using (a) the wearable imaging probe, and (b) the L11–5v probe. CPWC with 32 uniformly spaced steering angles was used in both cases. The common carotid artery (CCA), internal jugular vein (IJV), and vagus nerve (VN) are labeled.

B. Closed-Loop Optimization for Target Modulation

After confirming the imaging performance of the custom imaging probe and the beam steering and focusing capabilities of the custom modulation probe in Section III, we studied the closed-loop operation of the proposed procedure for image-guided neuromodulation (Fig. 11). This method addresses two of the key practical problems with image-guided neuromodulation, namely i) probe positioning and alignment errors, and ii) determination of the optimum neuromodulation parameters.

Fig. 11.

Flowchart of the proposed closed-loop image-guided ultrasound neuromodulation algorithm.

During probe positioning and alignment, a built-in strain sensor is used to estimate the curvature of the flexible PCB wrapped around the curved body surface (or ultrasound phantom). Next, suitable anatomical landmarks, if available, are first detected and then localized in the resulting B-mode images by using different approaches to estimate the position of the neuromodulation target. We use i) a convolutional neural network (CNN)-based image classification algorithm to detect the existence of the anatomical landmarks (in the case of the phantom, a tube that models a blood vessel) with high accuracy; and ii) a template matching algorithm to localize them in those images where landmarks are detected. The template matching algorithm is easier to implement with a limited amount of image data than higher-performance algorithms based on CNNs [49]. The system provides feedback to the user to move the probe if the algorithm fails to locate the target, and this human-in-the-loop process is continued until successful target detection. Thus, the method is only suitable for nerves located near suitable landmarks; other approaches are required if such landmarks are not available.

The template matching algorithm is based on storing every object to be recognized as a “template” in long-term memory [50]. For this purpose, the raw B-mode images are first filtered using histogram equalization and edge enhancement to improve contrast. Next, area-based template matching is performed. A standard region of interest (ROI), which is usually smaller than the target image, is selected from the template. The existence of the target at a given location ($u,v$) is then determined by comparing the normalized cross-correlation factor $y(u,v)$ between the image and the template ROI with a predetermined threshold. The former is defined as

$$y(u,v)=\frac{\sum _{x,y}\tilde{f}(x,y)\tilde{t}(x,y,u,v)}{{\left[\sum _{x,y}{\tilde{f}}^2(x,y)\sum _{x,y}{\tilde{t}}^2(x,y,u,v)\right]}^{\frac{1}{2}}},$$ (1)

where $f(x,y)$ is the image, $\tilde{f}(x,y)\equiv f(x,y)-{\bar{f}}_{u,v},{\bar{f}}_{u,v}$ is the mean value of $f(x,y)$ in the region of the image within the template, $\tilde{t}(x,y,u,v)\equiv t(x-u,y-v)-\bar{t}$ and $\bar{t}$ is the mean value of the template. Note that normalizing the cross-correlation factor ensures that the algorithm is robust to brightness changes, i.e., variations in image intensity. The same algorithm can also be used to detect the orientation of the target and then provide feedback to the user on correct probe orientation. The location of the target nerve is then estimated based on the distances between the nerve and anatomical landmarks (PTA and PTV) in a ground truth B-mode image, such as those shown in Fig. 10. The ultrasound beam generated by the modulation array is then steered and focused to the target location to implement neuromodulation. The optimal neuromodulation parameters (transmit power $V_{HV}$, center frequency $f_0$, pulse repetition frequency $f_{\mathit{\text{rep}}}$, etc.) are determined by closed-loop functional feedback, i.e., by real-time sensing of suitable functional signals, ${\mathit{\text{Sens}}}_{\mathit{\text{real}}}$, from the target and comparing them to user-set reference values, ${\mathit{\text{Sens}}}_{\mathit{\text{user}}}$.

V. Closed-Loop Acoustic Neuromodulation

In this section, we demonstrate the feasibility of the functional feedback paradigm by implementing an in vitro prototype of a closed-loop neuromodulation system. The system is tested on a custom active phantom which includes a custom-made embedded PZT sensor and an actuator that serve as an artificial neuronal firing model and a muscle twitching model, respectively. The average pulse rate of the surface motion signatures generated by the actuator (which mimics the motion of skeletal muscles innervated by the target nerve) serves as a functional feedback signal during closed-loop operation.

The dashed box in Fig. 11 summarizes the operation of the proposed system for emulating closed-loop neuromodulation. The procedure can be described in more detail as follows. Template matching is used to estimate the position of the target, and thus the focal point for the modulation array that results in peak acoustic pressure at the nerve location. Next, delay values are calculated for each modulation array element to achieve beam focusing at the target location and then implemented using the Verasonics transmitter platform. The pressure generated by the focused modulation beam is sensed by the custom-made embedded sensor inside the phantom to trigger the activation (aka neuronal firing) of an artificial nerve model based on a pre-defined threshold value. This process emulates the start of neuronal firing whenever the acoustic pressure exceeds the firing threshold of a target nerve. Motion signatures are sensed on the surface of the phantom to estimate the average rate of instantaneous muscle twitches (i.e., the artificial neural firing rate), which is then fed back to the system input to set the transmit voltage of the modulation array in real time and thus optimize the beam intensity. Further details of this process are presented in the next few sub-sections.

A. Functional Feedback Mechanisms

Effective acoustic neuromodulation requires closed-loop feedback to automatically optimize the modulation beam intensity at the targeted position. A functional feedback mechanism, which is usually related to physiological changes in the target, is thus a fundamental component for effective delivery of FUS therapy. However, in order for functional feedback to be feasible, the system must accurately deliver the necessary amount of acoustic energy to the target with acceptable spatial resolution. The proposed modulation array is capable of beam focusing and steering in 2D (along the $x$ and $z$ axes) for delivering peak acoustic pressure at the estimated position. However, poorly-controlled factors including the geometrical variance of the body-worn device among various patients and body part, imaging artifacts, and target localization errors make it difficult for an open-loop system to deliver adequate performance in real-world scenarios.

B. Active Phantom Design and Fabrication

The proposed neuromodulation system needs to pass all relevant safety concerns before being approved for testing on humans. Therefore, it is vital to perform initial in vitro measurements on an ultrasound phantom that simulates the relevant anatomical landmarks identified in the template matching algorithm (i.e. blood vessels) of the human body to accurately evaluate the proposed neuromodulation system. As explained earlier, the vagus and tibial nerves are two targets of interest for FUS neuromodulation. Here, we choose to design and fabricate an artificial phantom that can simulate the human ankle for tibial nerve modulation tests on a curved surface. Fig. 12(a) summarizes the anatomy of the human ankle, including veins, muscles, and nerves in the region. B-mode images of this region (captured from the right ankle of a human subject) using the wearable and L11–5v probes are shown in Fig. 12(b) and 12(c). The major blood vessels, namely the posterior tibial artery (PTA) and posterior tibial vein (PTV), are identifiable near the tibial nerve in the B-mode image acquired from the wearable imaging probe. As shown in Fig. 12(a), the posterior tibial nerve (PTN) is located near a vascular bundle that includes the PTA and PTV. Thus, these blood vessels can be used as anatomical landmarks to locate the tibial nerve. The nerve localization algorithm identifies the relative positions of the nerve and its landmarks (artery and vein) by measuring the distances between their center points in a ground truth image, as shown in Fig. 13(a) [51]. Given a fixed probe orientation and corrected anatomic landmark positions, the nerve in a new B-mode image acquired from the custom probe can then be estimated by finding the intersection of two circles, as shown in Fig. 13(b). Note that such a simple model suffices since the human-in-the-loop process (discussed in previous sections) is assumed to correct any probe orientation or positioning errors before nerve localization begins.

Fig. 12.

(a) Drawing of an axial section through the right ankle, with major structures labeled. Key acronyms: PTA/PTV: posterior tibial artery/vein, MM: medial malleolus (the others are not listed here for brevity). The blue arrow represents the typical hand-held ultrasound probe orientation used for locating the posterior tibial nerve (PTN). (b)-(c) B-mode images of the right ankle of a human subject using (b) the wearable imaging probe, and (c) the L11–5v probe. CPWC with 32 uniformly spaced steering angles was used in both cases.

Fig. 13.

Estimation of tibial nerve location based on the relative positions of the anatomical landmarks in a B-mode image. (a) A ground truth B-mode ankle image [51] with the anatomical landmarks (PTA and PTV). (b) Estimation of the tibial nerve position as the intersection between circles with radii equal to the artery-nerve and vein-nerve distances.

Fig. 14(a) illustrates the design of the proposed active phantom with built-in sensing and actuation mechanisms. We designed the phantom to contain fluid-filled (water) silicone tubes embedded within the inner layer to simulate the PTA and PTV within the ankle region. To simplify the design, the geometry of the ankle was approximated as a cylinder with a diameter of 12 cm, in agreement with the average value for adults. The diameter of the PTA ranges from 4.3 mm to 7.7 mm, while the PTV diameter can be estimated to be around 5 to 10 mm. We chose the tubing dimensions for the phantom accordingly. A third solid rod was embedded between the other two to represent the target, i.e., the PTN. We used a speckle-free flexible and soft silicone rubber material rather than perfect soft tissue mimicking materials as used in clinical calibration phantoms for filling the phantom in order to mimic the properties of human tissue.

Fig. 14.

Design of an active ultrasound phantom for emulating neuromodulation. Side views of the phantom: (a) design, and (b) photograph after fabrication with a custom-made embedded PZT sensor. (c) Positioning of the wearable probe on the curved surface of the active phantom. (d) Locations of the imaging and modulation arrays with respect to the embedded structures within the phantom.

C. Target Localization Within the Active Phantom

Imaging and neuromodulation were implemented by wrapping the probes around the fabricated active phantom, as illustrated in Fig. 14(b). The positions of the imaging and modulation arrays with respect to the embedded tubes and PZT sensors are also shown in Fig. 14(c) and 14(d). When wrapped around the human body (e.g. huamn ankle), the two probes, with a short distance of ~2.5 cm between them, automatically align along the longitudinal axis of the nerve. The focal points can then be steered to the correct position based on real-time target localization. Fig. 15 shows an example of successful target localization using template matching from a B-mode image of the phantom acquired by the 4 MHz custom imaging probe. The cross-correlation between a stored template of the target (the tube filled with water that models the PTA) and a newly acquired image is maximized near the center of the target region, as expected.

Fig. 15.

Imaging of the active phantom using the custom wearable probe, which was wrapped around the curved surface as shown in Fig. 14. (a) B-mode image of the phantom, including water-filled tubes and the PZT pressure sensor. (b) Spatial distribution of the normalized cross-correlation factor, $y(u,v)$, found by the template matching algorithm.

Target detection rates were estimated by acquiring 300 B-mode images of the active phantom at different angles and positions using the custom imaging probe, thus simulating the effects of probe orientation and positioning errors. Different amounts of random noise with a Rice distribution were then added to set the SNR of the magnitude images. Two target detection methods, namely i) a CNN-based classifier, and ii) a template matching algorithm, were then compared as a function of SNR. The chosen CNN, which was AlexNet [52], was trained using transfer learning to reduce the size of the required training data set. The results in Fig. 16(a) indicate that the CNN-based classification algorithm greatly outperforms template matching in terms of detection accuracy. However, template matching is more convenient to use than the CNN-based algorithm, which requires the effort of preparing a sufficient amount of training data.

Fig. 16.

(a) Target detection accuracy as a function of the image SNR when using i) AlexNet-based image classification, and ii) a template matching algorithm. (b) Target localization error as a function of the image SNR when using the template matching algorithm.

To locate the anatomical landmarks (here, a tube modeling a blood vessel), we selected a template-matching algorithm instead of higher-performance object detection or image segmentation algorithms. This decision was primarily based on the lower effort required to collect and label large quantities of image data. The target localization error was computed as the average Euclidean distance to the center point of the predicted target location in Fig. 16(b) for all 150 images. The results are small enough to successfully detect the landmarks for SNR >12 dB, with a mean and variance <1.15 mm and <3.27 mm, respectively. Anatomical landmarks of live subjects can be detected with minimal inter-subject variability by first using the proposed human-in-the-loop process to ensure that the wearable probe is correctly oriented and that the target is centered in the imaging window.

One challenge in optimizing the modulation beam profile for maximal acoustic pressure at a target location is to develop a method to measure the pressure amplitude at that location. Experiments showed that temperature sensing is ineffective at the relatively low intensities of interest. Thus, we developed an active phantom with an embedded piezoelectric pressure sensor to directly measure the pressure at the targeted region and to serve as a receiver of neural stimulus before propagating nerve impulses to generate muscle activation (twitches). For this purpose, a PZT plate with dimensions of 1 cm and 4 cm and a thickness mode resonance frequency of ~1.3 MHz (to match the frequency of the modulation array, thus maximizing sensitivity) was attached to a solid rod embedded inside the phantom. The sensor is aligned and placed 1.5–2 cm deep inside the phantom between the two silicone tubes to face the modulation array at the estimated position of the target nerve.

The signal and ground electrodes of the PZT plate were soldered to wires that allow the sensor to be connected to data acquisition equipment (e.g., an oscilloscope) or to an artificial neural firing model during closed-loop neuromodulation phantom experiments. Fig. 14(c) and 14(d) illustrate the geometry of the fabricated active phantom, including the embedded fluid-filled tubes and PZT sensor.

VI. Experimental Results

This section first describes the experimental setup used to evaluate a functional feedback system for closed-loop acoustic neuromodulation using the active phantom introduced in the previous section. Experimental results are then demonstrated to validate the performance of our system in optimizing the transmit voltage and the resulting acoustic pressure based on the chosen functional feedback signal (muscle motion).

A. Experimental Setup

We designed and implemented a hardware prototype for emulating the closed-loop functional feedback mechanism during non-invasive real-time low-power acoustic neuromodulation of peripheral nerves, for example the PTN. An overview of the closed-loop system is shown in Fig. 17(a). The design is based on optimizing the transmit voltage, $V_{HV}$, and the acoustic pressure generated by the modulation array based on a functional feedback signal, namely real-time estimation of the average muscle twitching rate. For this purpose, an inertial measurement unit (IMU) is used to sense the motion signatures generated on the surface of the phantom when the emulated neuron fires. Muscle twitching and neural firing are both modelled artificially and actuated through an FUS stimulator. The average rate of emulated muscle twitches is estimated in real-time from the instantaneous neuronal firing activity, which follows a Poisson process. The rate is controlled and regulated automatically by a proportional and integral (PI) controller that sets $V_{HV}$ for the ultrasound transmitter to drive FUS stimulation. The latter is computed for each time window based on a firing rate error signal, i.e., by comparing the estimated average muscle twitching rate, ${\bar{r}}_{\mathit{\text{sens}}}$, with a user-set rate, $r_{\mathit{\text{set}}}$. Thus, the system emulates the modulation of nerve activity and associated muscle contractions (i.e., twitches) when image-guided FUS is applied in vivo.

Fig. 17.

Functional feedback mechanism. (a) High-level view of the feedback loop, and (b) a detailed hardware block diagram.

The closed-loop system in Fig.17(a) is implemented using five main hardware components, as summarized in the illustrative block diagram of Fig. 17(b) and the experimental setup in Fig. 18(a): the active ankle phantom, an ultrasound transmitter (Verasonics research ultrasound system), two microcontrollers (MCU1 and MCU2) on Teensy 4.0 boards, an air pump, a valve, a valve controller, and an analog front-end (AFE). Transmit pulses were generated by the Verasonics ultrasound system shown in Fig. 18(a) to excite both the imaging and modulation arrays. The driving voltage, $V_{HV}$, required to generate adequate acoustic intensity was configured by a custom MATLAB program (denoted by “Controller” in Fig. 17(b)) based on the output of a PI controller. The controller regulates the average muscle twitching rate, as estimated from motion signatures sensed on the surface of the phantom.

Fig. 18.

(a) Experimental setup used for demonstrating closed-loop functional feedback for ultrasound neuromodulation. (b) Photograph of the fabricated wearable probe, which integrates the modulation and imaging arrays on a flexible PCB substrate.

The transmit pulses generated by the Verasonics platform excited the custom modulation probe to implement FUS in real-time. During FUS experiments, the probe was positioned right above the neuromodulation target (i.e., the PZT sensor and solid rod) to maximize the acoustic pressure available for activating the neuronal firing model with the flexible probe. For this purpose, the probe (shown in Fig. 18(b)) was wrapped tightly around the phantom, as shown in Fig. 14(b) and 14(c). If the location of the neuromodulation is unknown, the imaging probe on the flexible PCB can be used for target localization based on nearby anatomical landmarks (such as blood vessels), as described in the previous section.

An analog front-end (AFE) implemented on a custom PCB was used to receive, identify and further process the PZT sensor output signal, $S_{\mathit{\text{sens}}}$, for neuronal firing activation. The AFE first filters out noise outside the bandwidth of the modulation waveform and then amplifies the filtered signal via a variable-gain amplifier (VGA) realized using an op-amp (OPA357, Texas Instruments). The amplified pressure signal is mapped to the mean rate of the Poisson process that models neural firing. For this purpose, a diode-based envelope detector (ED) within the AFE board was used to extract the peak amplitude, $s_p$, of the pressure pulses after low-pass filtering with a cut-off frequency of 16 kHz. The latter was then sampled every $5\mu \text{s}$ and digitized by an ADC in MCU1 to obtain a number, $S_n$, and then converted to the mean firing rate, $r_n$, via a look-up table. The look-up table contains a list of $S_n$ and $r_n$ values which were produced based on a sigmoid function that emulates the nature of real-life neural firing rates. For this purpose, $s_p$ values were measured by sweeping $V_{HV}$ from 0 V to $100{\text{V}}_{pp}$ with a uniform step size of 5 V and linearly interpolated to cover the entire $V_{HV}$ range with a finer resolution of 0.01 V. The value of $r_n$ was configured from 0 Hz to 5 Hz to cope with the slow response time of the valve that triggers the generation of air pulses (i.e., emulated muscle twitches).

To emulate neural activation based on an acoustic pressure threshold, a comparator circuit was built to generate an artificial neural firing trigger signal, $S_{\mathit{\text{trig}}}$, whenever $s_p$ exceeds the preset neuronal firing activation threshold, $V_{th}$. A program running in MCU1 is triggered by $S_{\mathit{\text{trig}}}$ to deliver the instantaneous neuronal firing trigger or valve control signal, $S_{\mathit{\text{fire}}}$. These emulated neural spikes are randomly generated by a Poisson process model. They drive a power MOSFET switch (IPA093N, Infineon) that emulates motor excitation in the nervous system by controlling a valve that sets the air flow rate within a silicone tube embedded in the phantom.

The valve mentioned above is connected between an air pump and the silicone tube, as shown in Fig. 14(a). Note that in this experiment, the tube was filled with air instead of water to efficiently mimic muscle twitches due to the large expansion of air within the tube. Thus, emulated muscle twitches, in the form of air pulses, are generated whenever the valve opens. Both the artificial tibial nerve model and the emulated muscle are located at a depth of ~1.5 cm below the surface of the phantom, in agreement with the typical depth of the PTN and the skeletal muscles that it innervates. The air pulses flowing along the tube within the phantom cause its surface to vibrate in sync with the air pressure waveform, thus emulating the effect of muscle twitches.

The Poisson process that generates neural spikes was driven by $r_n$ and a small time interval, $dt$. The instantaneous neuronal firing signals are generated by the algorithm below:

$$S_{\mathit{\text{fire}}}\left(x_{\mathit{\text{rand}}}\right)=\left\{\begin{array}{ll}1,& \text{if}{x}_{\mathit{\text{rand}}}<r_n\cdot \Delta t\\ 0,& \text{otherwise},\end{array}\right.$$ (2)

where $x_{\mathit{\text{rand}}}$ is a random number with a uniform distribution between 0 and 1 and $S_{\mathit{\text{fire}}}\left(x_{\mathit{\text{rand}}}\right)$ defines whether the neuron fires during the time step of duration $\Delta t$.

The vibration motion signatures, $S_{vib}$, on the surface of the phantom were sensed using a digital motion sensor at a sampling frequency of 20 kHz installed on a flexible PCB wrapped around the phantom next to the wearable probe. The sensor is a six-axis inertial motion unit (IMU) (ICM42605, TDK InvenSense) and can be easily implemented on the wearable probe if required. MCU2 was used to collect motion signatures, $S_{vib}$, from the three-axis accelerometer within the IMU. The number of signatures with amplitudes exceeding a preset threshold was repeatedly counted to estimate the number of pulse peaks (i.e., neural spikes) within a fixed 5 sec window. The threshold was selected to be low enough to detect the peaks of pulses generated over the entire range of transmit voltages. Since the amplitudes of the motion signatures are unknown in real-life, a threshold of 3 to 5 times the standard deviation of the data in the previous time window can be used for peak detection. The average rate of muscle twitching, ${\bar{r}}_{\mathit{\text{sens}}}$, was then estimated for each pulse window and compared with the user-set rate $r_{\mathit{\text{set}}}$ to generate the error signal, $ϵ$. The latter serves as the input of the PI controller which automatically sets the transmit voltage, $V_{HV}$, thus completing the closed-loop system.

One source of counting error is the fact that some pulses can have multiple peaks above the threshold. To minimize the probability of multiple peak counts during a single pulse, each motion signature is counted only when the pulse crosses the threshold for the first time, after which counting is disabled for a dead time of 175 ms. The value of dead time is chosen to be slightly shorter than the smallest inter-spike interval (approximately 200 ms). The resulting update rate of the counter corresponds to a time step of $\Delta t=200\text{ms}$. The counter is reset and restarted at intervals of $t_w$, the duration of the averaging window. The average neuronal firing rate is then estimated as

$$\bar{r}=\frac{N_{\mathit{\text{peak}}}}{t_w}=\frac{N_{\mathit{\text{peak}}}}{N_{\Delta t}\cdot \Delta t},$$ (3)

where $N_{\mathit{\text{peak}}}$ and $N_{\Delta t}$ are the number of pulse peaks and the number of time steps within $t_w$, respectively.

B. Experimental Results

Prior to experimentally assessing the functionality of the closed-loop feedback system on the active phantom in Fig. 17, the peak pressure amplitude, $S_p$, was measured in terms of voltage, $V_p$, across a frequency range of 0.7 MHz to 2 MHz and a $V_{HV}$ range of 10 V to 50 V at the envelope detector output to characterize the sensor's sensing ability in the frequency domain and to estimate the sensitivity of the PZT pressure sensor. The data points of the sensor sensitivity plot were fitted by a linear function of the form

$$V_p=a_1{V}_{HV}+b_1,$$ (4)

where $a_1$ and $b_1$ are constants.

The measured frequency response of the modulation array when wrapped around the phantom using the embedded sensor is shown in Fig. 19(a), in agreement with the response measured using a hydrophone in water demonstrated in Fig. 5(a). The measured sensitivity curve and its linear fit, which has slope and offset parameters of $a_1=0.0203$ and $b_1=-0.0159$, respectively, are shown in Fig. 19(b). The slope, $a_1$, defines the transfer function, $G_2\left(s\right)$, of the linear subsystem where the driving voltage, $V_{HV}$, is an input and the peak pressure voltage, $V_p$, is an output. The fitted value of $V_p$ was then mapped to an expected artificial neural firing rate, $r_n$, via the sigmoid function shown in Fig. 19(c). The latter approximates the expected effects of ultrasound neuromodulation, i.e., the dependence of average artificial neural firing rate on peak pressure. Note that the response is assumed to be excitatory, but can be readily extended to the inhibitory case by reversing the sign of the sigmoid function. In either case, the linear region of the function can be approximated by a linear function of the form

$$r_n=a_2{V}_p+b_2,$$ (5)

where $a_2$ and $b_2$ are constants.

Fig. 19.

(a) Measured peak envelope voltage, $S_p$, of the modulation array using the embedded PZT sensor as a function of frequency. (b) Measured peak envelope voltage, $S_p$, of the embedded PZT sensor output, which is proportional to the peak pressure, as a function of the driving voltage, $V_{HV}$. (c) Mapping of the artificial neural firing rate to the peak pressure signal shown in (c) by using a sigmoid function that approximates the behavior of real-life neurons. (d) Instantaneous IMU output measured within a 5 s window that contains several emulated neural spikes. (e) Motion signature of a single neural spike measured by the IMU. (f) Output of the functional close-loop feedback system when the input frequency changes from 0 Hz to three different user-set $r_{\mathit{\text{set}}}$ values, namely 3.5 Hz, 2 Hz and 3 Hz. (g) Rise time and variance of steady-state error for the step response as function of the integral parameter ($K_i$) when the proportional parameter ($K_p$) is fixed at 0.5. (h) Rise time and variance of steady-state error for the step response as a function of the proportional control parameter ($K_p$) when the integral parameter ($K_i$) is fixed at 0.5.

The proposed mapping of peak pressure to firing rate is plotted in Fig. 19(c). The slope, $a_2$, of the linear region (which has a value of 9.395) then gives an approximation of the transfer function, $G_3\left(s\right)$, of the peak pressure-firing rate mapping system. Including the transfer function of the PI controller, $G_1\left(s\right)=K_p+\frac{K_i}{s}$ where $K_p$ and $K_i$ are the proportional and integral control terms, respectively, the feedforward transfer function of the closed-loop feedback system in Fig. 17(a) can be expressed as

$$G\left(s\right)=G_1\left(s\right)G_2\left(s\right)G_3\left(s\right)=a_1{a}_2\left(K_p+\frac{K_i}{s}\right).$$ (6)

The signal averaging operation can be approximated by a low-pass filtering sinc function, $\text{sinc}\left(s{t}_w/2\right)$, followed by a time delay of $\frac{-t_w}{2}$ where $t_w=5\text{s}$ is the duration of the spike train averaging window. An additional time delay of $t_d\approx 2\text{s}$ is required to search the look-up table that sets the firing rate. Thus, the overall transfer function in the feedback path of the system can be estimated as

$$H\left(s\right)=\text{sinc}\left(s{t}_w/2\right)e^{\frac{-t_{w}s}{2}}{e}^{-t_{d}s}$$ (7)

The final closed-loop transfer function of the emulation system can then be expressed as

$$\frac{\bar{Y}\left(s\right)}{X\left(s\right)}=\frac{G\left(s\right)H\left(s\right)}{1+G\left(s\right)H\left(s\right)},$$ (8)

where $Y\left(s\right)$ is the instantaneous artificial neural firing rate signal and $\bar{Y}\left(s\right)=H\left(s\right)Y\left(s\right)$ is the averaged version of $Y\left(s\right)$.

During experimental verification of closed-loop acoustic neuromodulation using the phantom, the muscle activation threshold, $V_{th}$, for generating air pressure pulses was set at a relatively low value of 0.2 V, while the voltage gain of the VGA within the AFE was configured as ~4.7.

Fig. 19(d) shows a typical instantaneous IMU output over a 5 s window. Motion signatures corresponding to several air pulses (i.e., muscle twitches) are visible. The waveform of a single pulse is shown in more detail in Fig. 19(e). The detection threshold used to count pulse peaks for estimating the average artificial neural firing rate was set at 0.3 V, which results in a good trade-off between true and false positive rates. The estimated artificial neural firing rate in this example, as obtained from the pulse peak count within the 5 s time window visible in Fig. 19(d), was 1.6 Hz. This value is close to the input firing rate $r_n$ of 1.9 Hz used to generate neural spikes, i.e., the average rate of the underlying Poisson process. More quantitatively, the variance of $N_{\mathit{\text{peak}}}$, the number of pulses occurring within a time window of length $t_w$, is equal to its mean value, ${\bar{N}}_{\mathit{\text{peak}}}$, as for any Poisson process. Thus, the signal-to-noise ratio (SNR) of the firing rate estimate, ${\bar{r}}_{\mathit{\text{sens}}}$, is equal to ${\bar{N}}_{\mathit{\text{peak}}}^2/{\bar{N}}_{\mathit{\text{peak}}}={\bar{N}}_{\mathit{\text{peak}}}$ in power units. The chosen value of $t_w=5\text{s}$ ensures that ${\bar{N}}_{\mathit{\text{peak}}}>5$ for typical firing rates (>1 Hz), i.e., results in SNR values > 5. The precision of the firing rate estimate can be further improved by using longer averaging windows, i.e., increasing the value of $t_w$. However, this also increases the time delay, $t_d$, within the loop transmission, as shown in Eqn. (8), which has a negative impact on loop stability. If required, this trade-off between estimation accuracy and loop stability can be somewhat relaxed by using overlapping windows.

Fig. 19(f), 19(g) and 19(h) show typical experimental results obtained during closed-loop operation. Fig. 19(f) verifies that the closed-loop system can automatically regulate the average rate of neural activation (i.e., muscle twitching) by using the PI controller to adjust the drive voltage that generates the corresponding acoustic pressure. The effects of the PI controller parameters ($K_p$ and $K_i$) on the rise time and the variance of the steady-state error for the closed-loop system step response are also evaluated separately when $K_p$ and $K_i$ are set to fixed values of 0.5 and 0.3, respectively. The results of Fig. 19(g) and 19(h) both illustrate a trend of decreased rise time and increased steady-state error variance with the value of $K_p$ or $K_i$, respectively. These trends arise from the fact that the closed-loop system bandwidth is an increasing function of both $K_p$ and $K_i$. Thus, they match the expected effects of the two controller parameters on system operation.

The results described above show that our closed-loop acoustic neuromodulation system can adaptively determine the transmit voltage, $V_{HV}$, required to drive the modulation array probe. The proposed PI controller implements feedback control to set $V_{HV}$ above the activation threshold for artificial neural firing by sensing and regulating the average muscle twitching rate. The latter is estimated from surface motion signatures (vibration patterns) sensed by the IMU. The maximum value of $V_{HV}$ is currently limited to $100{\text{V}}_{pp}$ by the Verasonics research ultrasound system used during the experiments. However, experimental verification of neuromodulation on an animal or human body in vivo would benefit from increased values of $V_{HV}$ (say, as high as $100{\text{V}}_{pp}$). The increased transmit voltage, which can be generated by a custom portable ultrasound transceiver, would ensure that sufficient acoustic pressure can be generated to modulate neural activity and muscle activation rates while allowing the system to compensate for variations in experimental parameters such as probe position/orientation, tissue properties, and inter-subject variability.

C. Discussion

The proposed closed-loop system includes all major components of a wearable and autonomous platform for non-invasive and effective FUS neuromodulation on an active phantom. The effectiveness of the proposed method was verified through a novel bench-top setup that emulates human muscle activation, as described in this section. Table III compares our system with three of the recent notable studies on flexible ultrasound array designs [25], [35]) and optimized FUS neuromodulation through functional feedback [9]. The work by Seok et al. [53] suggests a wearable neuromodulation device using a 1D capacitive micromachined ultrasonic transducer (CMUT) array controlled by a front-end application-specific integrated circuit (ASIC). However, neuromodulation through the device was not image-guided for the purpose of target localization. Also, this work did not explore the functional feedback mechanisms for autonomous neuromodulation applications. The second work by Li et al. [54] presented an image-guided dual-mode neuromdoulation probe for multi-target applications. This device is capable of B-mode imaging at high frequencies, but the bulky and non-wearable transducer setup makes it difficult to use on live subjects. The third work by Konofagou et al. [9] did present a functional feedback mechanism for neuromodulation. However, the system used typical commercial rigid probes, which limits its use and effectiveness for closed loop autonomous applications.

TABLE III.

Comparison With Earlier Work

	Transducer Fabrication Method	Transducer Bonding Method	Wearable (on Curved Surfaces)	B-Mode Imaging	Functional Feedback
Seok et al. [53]	Micromachining	Wire-bonding	✓	N/A	N/A
Li et al. [54]	N/A	N/A	N/A	✓	N/A
Konofagou et al. [9]	N/A	N/A	N/A	✓	✓
This work	Dice-and-fill 2-2 composite	Low viscosity epoxy attachment	✓	✓	✓

By contrast, our work realizes a wearable neuromodulation platform featuring functional feedback mechanism for optimized and autonomous operation on a phantom. Our fabrication method combines the low-cost 2–2 composite dice-and-fill fabrication process with the low viscosity epoxy attachment assembly process, thus enabling a high array fabrication throughput with good imaging and beam focusing quality. Unlike earlier work, our system is capable of generating B-mode images with the spatial resolution and CNR comparable to commercial rigid arrays on both human tissues like neck and ankle blood vessels and curved/conformal surfaces that can be readily analyzed using template matching or other methods (e.g., CNNs) to localize the target position. A functional feedback control algorithm was implemented for proof of functionality. Future work can use alternate functional feedback signals and/or directly estimate the pressure distribution near the target from B-mode images (e.g., by sensing displacement-induced phase shifts) for more accurate control of neuromodulation.

VII. Conclusion

We have presented the successful implementation of a closed-loop approach using a functional feedback mechanism for low-intensity focused ultrasound modulation of peripheral nerve activity on a custom active phantom. Our approach is based on using a wearable dual-mode ultrasound probe containing piezoelectric transducer arrays installed on a flexible PCB to implement imaging and neuromodulation, respectively. B-mode images with comparable resolution and CNR to commercial rigid probes at a higher frequency are generated by the wearable imaging arrays (on a calibration phantom, human tissues and curved surfaces) and analyzed using a template matching algorithm to detect the location of anatomical landmarks (blood vessels) near the targeted nerve. This information is then used to direct the modulation beam towards the target. The effects of probe positioning errors, reflections at the probe-skin boundary, inter-subject variability, and other unknown factors on the pressure delivered to the target are compensated in real-time by using a built-in motion sensor to monitor and estimate functional feedback signals, namely the average rate of instantaneous muscle twitching on the skin surface, and adjust the transmit voltage appropriately. Muscle twitching is modelled through a Poisson process and the expected average rate is mapped to the acoustic pressure using a sigmoid function to simulate the nature of neuronal firing with saturation at the upper and lower pressure limits. The proposed concept was demonstrated on a benchtop prototype by using an active tissue phantom that includes embedded models of the posterior tibial nerve, nearby blood vessels, and innervated skeletal muscle.

Biographies

Junjun Huan received the B.S. degree in electronic information engineering and the M.S. degree in electrical engineering from the University of Dayton, Dayton, OH, USA. He is currently working toward the Ph.D. degree in electrical and computer engineering with the University of Florida, Gainesville, FL, USA. His research interests include electronics for ultrasound imaging and image-guided treatment, wearable medical sensors, and hardware security.

Vida Pashaei received the B.Sc. and M.Sc. degrees in electrical engineering from the University of Tabriz, Tabriz, Iran, in 2010 and 2012, respectively, and the Ph.D. degree in electrical engineering from the Case Western Reserve University, Cleveland, OH, USA, in 2021. She was a Visiting Researcher with the University of Florida from 2019 to 2021. She is currently a Sensor Design Engineer with Apple Inc., San Diego, CA, USA. Her research interests include sensor and actuator design, wearable electronics, biomedical circuits, and systems with a special focus on ultrasound imaging and neuromodulation, image processing, and precision instrumentation for sensor interface applications.

Steve J. A. Majerus (Senior Member, IEEE) received the B.S., M.S., and Ph.D. degrees in electrical engineering from the Case Western Reserve University (CWRU), Cleveland, OH, USA. His primary training was in ultra-low-power application-specific integrated circuit (ASIC) design for miniaturized implantable medical devices, and he has developed ASICs for pressure sensing, neural recording, and wideband intravascular ultrasonic imaging. He has additional experience developing ASICs (in Si, SoI, and SiC) for high-temperature and aerospace applications. His doctoral thesis focused on developing and demonstrating the first cystoscopically-implanted, catheter-free bladder pressure monitor.

Swarup Bhunia (Fellow, IEEE) received the B.E. (Hons.) from Jadavpur University, Kolkata, India, the M.Tech. degree from the Indian Institute of Technology (IIT), Kharagpur, and the Ph.D. degree from Purdue University, IN, USA. Currently, he is a Professor and a Semmoto Endowed Chair with the University of Florida, FL, USA. Earlier he was appointed as the T. and A. Schroeder Associate Professor of Electrical Engineering and Computer Science with the Case Western Reserve University, Cleveland, OH, USA. He has over ten years of research and development experience with over 200 publications in peer-reviewed journals and premier conferences. His research interests include hardware security and trust, adaptive nanocomputing and novel test methodologies. He received the IBM Faculty Award (2013), the National Science Foundation CAREER Award (2011), the Semiconductor Research Corporation Inventor Recognition Award (2009), and the SRC Technical Excellence Award as a Team Member (2005), and several best paper awards/nominations. He has been serving as an Associate Editor of IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, IEEE Transactions on Multi-Scale Computing Systems, ACM Journal of Emerging Technologies, served as a Guest Editor of IEEE DESIGN & Test of Computers (2010, 2013) and IEEE Journal on Emerging and Selected Topics in Circuits and Systems (2014). He has served in the organizing and program committee of many IEEE/ACM conferences.

Soumyajit Mandal (Senior Member, IEEE) received the B.Tech. degree from the Indian Institute of Technology (IIT) Kharagpur, India, in 2002, and the S.M. and Ph.D. degrees in electrical engineering from Massachusetts Institute of Technology (MIT), Cambridge, MA, USA, in 2004 and 2009, respectively. He was a Research Scientist with the SchlumbergerDoll Research, Cambridge (2010–2014), an Assistant Professor with the Department of Electrical Engineering and Computer Science, Case Western Reserve University, Cleveland, OH, USA (2014–2019), and an Associate Professor with the Department of Electrical and Computer Engineering, University of Florida, Gainesville, FL, USA (2019–2021). He is currently a Research Staff Member with the Instrumentation Division, Brookhaven National Laboratory, Upton, NY, USA. He has over 175 publications in peer-reviewed journals and conferences and has been awarded 26 patents. His research interests include analog and biological computation, magnetic resonance sensors, low-power analog and RF circuits, and precision instrumentation for various biomedical and sensor interface applications. He was a recipient of the President of India Gold Medal in 2002, the MIT Microsystems Technology Laboratories (MTL) Doctoral Dissertation Award in 2009, the T. Keith Glennan Fellowship in 2016, and the IIT Kharagpur Young Alumni Achiever Award in 2018.