Introduction

Ultrasound with a long penetration range, focusing capability, and well-defined safety guidelines is an attractive modality for medical diagnostics and therapeutics1,2,3,4,5,6,7,8. Thus, there have been tremendous advancements in terms of ultrasound image acquisition and processing over the past few decades9,10,11. However, despite these advancements, medical ultrasound hardware systems remain bulky and expensive and are only available in hospitals. Only recently, there have been movements to miniaturize these systems for portable and wearable applications, including Butterfly Networks' hand-held probes for smartphones and piezoelectric-based flexible arrays12,13,14,15,16. While portable probes still require a fixture or hand to position the probe around our body, an ultrasound patch can serve as a next-generation platform for point-of-care and home healthcare applications, as it is completely hands-free, free of body locations, and enables continuous and long-term diagnostics and treatment17,18,19,20,21.

One of the most essential components of patch-type ultrasound systems is the ultrasound transducer itself, which interfaces directly with the body. The ultrasound arrays should be flexible to conformally attach to our body, but still offer high acoustic power to achieve high image resolution and therapeutic efficacy. Moreover, to be able to disseminate the patch technology to every home, the flexible ultrasound transducers should be manufactured through mass production to minimize the cost. Recently, two strategies have been proposed to implement ultrasound patches: the use of completely rigid piezoelectric transducers with a soft coupling sticker22,23,24 and the development of completely flexible transducer arrays based on piezoelectric ceramics and silicone elastomer18,25,26,27. While both approaches offer a stable interface with the body and long-term interrogation for medical diagnostics, the design space of the transducers is still limited, and the transducers are not scalable for mass production due to manual processing.

In contrast to piezoelectric materials-based transducers, capacitive micromachined ultrasound transducer (CMUT) is a silicon-based technology that offers a high degree of design freedom, enables mass production, and allows for thin structures4,28,29,30,31. A few studies have reported flexible ultrasound transducers based on the CMUT technology32. To achieve flexibility, silicon membranes were replaced with polymers such as polyimide (PI) or polyethylene terephthalate (PET), which demonstrated operations in a bending state but suffered from low acoustic power because of low Young’s modulus of the vibrating membrane33,34,35. Another work on 2D CMUTs on a flexible substrate suffered from low transmission efficiency because the posts could not firmly clamp the edges of vibrational membranes due to the vibration of the substrate35. Most importantly, similar to the case of flexible PZT arrays, the movement of skin and body can readily change the direction of the ultrasound beam of each element, which is detrimental to accurate targeting for therapeutics and image acquisition for medical diagnostics.

Here, we report a flexible CMUT based on flex-to-rigid (FTR) technology, allowing for statically adjustable bending modulation using a low-melting-point metal alloy (LMPA) (Fig. 1a). The flex-to-rigid scheme incorporates rigid silicon membranes and substrates connected by flexible elastomer bridges, which enable higher acoustic power compared to conventional flexible CMUTs based on a soft substrate and membrane36 (Fig. 1b). By developing FTR technology for CMUTs at wafer-scale, the advantages of CMUT technology, including mass production and design flexibility, are maintained while endowing mechanical flexibility. In addition, through the use of LMPA, our structures can also maintain a specific bending curvature to ensure accurate and statically adjustable ultrasound beam forming. The phase of the LMPA (Melting point: 47.2 °C) embedded in our FTR structures can shift from solid to liquid and vice versa through Joule heating. This is the level at which no damage occurs to the skin, as the maximum skin temperature remains below 60 °C after 5 seconds of exposure, according to guidelines discussed by the Occupational Safety and Health Administration (OSHA) in the United States37. The curvature is not permanent but can be readily changed between inner and outer bending states (Fig. 1c). The ability to create a specific bending curvature without altering device characteristics allows for a single, streamlined operation and accurate positioning. Moreover, this advancement effectively focuses ultrasound to a specific distance without the need for beamforming.

Fig. 1: Design flexibility and application of flex-to-rigid (FTR) CMUT.
figure 1

a A Schematic diagram of anti-inflammatory using ultrasound neuromodulation of the spleen. b 3D illustration of the flex-to-rigid FTR CMUTs. c Principle of the dynamic bending modulation using a phase transition of low-melting-point metal alloy through joule heating. d The FTR structure consisting a flexible area that is etched in a trapezoidal shape and filled with an elastomer and a device area based on a hard Si substrate. e Photographs of Inner and outer bending with FTR structure composed of Si and PDMS elastomer (scale bar: 4 mm) f A photograph of fabricated FTR structures in various shapes (scale bar: 1 cm). g, h Flexible FTR structures with different sizes of rectangular rigid structures (scale bar: 1 mm). i, j Flexible 2D and 3D cubic shapes (scale bar: 5 mm). k, l Flexible 2D and 3D flower shapes (scale bar: 5 mm).

Full size image

To demonstrate the potential of FTR CMUT technology with bending modulation, we design and fabricate two FTR CMUTs with different center frequencies of 3.4 and 1.5 MHz in immersion. We develop 1D arrays to demonstrate the proof-of-concept of our technology for therapeutic applications. For the therapeutic application, we conduct ultrasound stimulation of the spleen in vivo, which alleviates autoimmune diseases by suppressing autoimmune cells such as T and B cells secreted from macrophages stimulation38,39,40,41,42. The 1D CMUT array provides sufficient coverage of the mouse spleen and curvature for the abnormal region. The electrical and acoustic characteristics of FTR CMUTs showed that the characteristics are maintained after 100 cycles of inner and outer bending (radius of curvature: 2.5 mm). Using 1D FTR CMUT arrays, we successfully demonstrated the potential of our device for long-term therapeutic applications by stimulating the spleen of mice and observing the alleviation of rheumatoid arthritis through behavioral experiments.

Results

Design and fabrication of flex-to-rigid (FTR) CMUT

CMUT arrays typically consist of multiple elements, and each element is composed of a number of electrically-shorted circular vibrating membranes (Supplementary Fig. 1). Each circular membrane is the basic unit of the CMUT, which generates ultrasound upon electrostatic actuation. The circular membrane is suspended over a vacuum cavity, and DC and AC voltages applied across the top and bottom electrodes separated by the vacuum cavity actuate the top circular membrane to generate ultrasound. A typical structure is composed of all rigid materials: single-crystal silicon for the membrane, silicon dioxide for the post, and silicon for the substrate. To provide flexibility, we designed the FTR structures such that each element that generated ultrasound remained in the typical rigid form while the spacings between the elements were filled with elastomer (Fig. 1d, e). Thus, the key development was to design spacings between the elements to provide maximum bending range and to create reliable spacings between the elements on a wafer scale without damaging the rigid CMUT structures.

We first devised a wafer-scale fabrication of FTR structures compatible with standard CMUT fabrication. A typical strategy of FTR fabrication is to first fabricate the devices on a silicon wafer and then perform through-wafer etching to expose the spacing between the devices43. The devices are typically held through thin but rigid suspended bridges. Our strategy was to eliminate the rigid bridges and fill these spacings with elastomer to endow flexibility. For the backside silicon etching, we adopted an anisotropic wet etching method of the Si{100} wafer, which utilizes the different etch rates of silicon crystal planes to expose a < 111> plane in a trapezoidal shape (Supplementary Fig. 2a)44. We chose tetramethylammonium hydroxide (TMAH) as the etchant because of its CMOS compatibility and high etch selectivity between silicon and silicon dioxide, which is often used as a hard mask (Supplementary Fig. 2b).

We fabricated the FTR structures without CMUT to confirm that our process provided accurate spacings between the devices and sufficient spacings for bending. The silicon dioxide (SiO2) layer was used as a hard mask to pattern the thick silicon substrate. The 450-µm-thick bottom silicon was etched through TMAH with Hydroxylamine (NH2OH) solution at 80 °C. For the mask opening areas of 400, 600, and 800 µm, the final opening at the device level (i.e., the spacing between CMUT elements) were 70.8, 277.0, and 412.6 µm, respectively (Supplementary Fig. 2c–h). The cross-sectional image of the FTR structure showed a trapezoid structure with an angle of 54.7° as expected. The trapezoidal gap between the elements endows high flexibility in both inner and outer deformation. At a 2 mm radius of curvature in inner and outer bending states, the FTR structure could be bent without damaging the rigid structure on top (Supplementary Fig. 3). In addition, various shapes of FTR structures such as planes, cubes, and closed flowers were fabricated to show the versatility of our fabrication (Fig. 1f–l), which shares conceptual similarities with previously reported flexible patch designs using rigid-island and soft-interconnect structures45,46,47. The cube and flower designs demonstrate the potential of our FTR fabrication technology, which can provide conformal and statically adjustable coverage of rigid acoustic devices on not only 2D surfaces but also on various shapes of 3D targets.

Next, we fabricated a flexible CMUT based on the developed FTR fabrication (Fig. 2a, Supplementary Fig. 4). Prior to substrate etching to create the spacing, CMUT circular vacuum cavities were first formed using the standard CMUT wafer-bonding process, but with oxide fusion bonding. A high bonding yield was achieved throughout the wafer (Fig. 5)48. After back-side etching to create the spacing, the elements were defined, and top and bottom electrodes were deposited to operate CMUT. The LMPA was coated on the etched parts through a needle of a micro syringe, and the PDMS was molded on the backside of a wafer to enable dynamic bending modulation without leakage of LMPA (Fig. 2b). The LMPA is positioned on the backside of the CMUT, ensuring that there is no direct contact with the skin during operation. Additionally, the PDMS was used as a protective barrier to reduce the risk of skin exposure to LMPA. The FTR CMUT could be bent smoothly inwards and outwards without any cracks or delamination of layers on the CMUT elements (Fig. 2c–e, Supplementary Fig. 6). The SEM cross-sectional image shows that the electrode (100 nm), membrane (450 nm), vacuum cavity (150 nm), and insulation layer (100 nm) were well-defined with dimensions matching the mask design (Fig. 2f).

Fig. 2: Fabrication of the FTR CMUTs.
figure 2

a 3D illustration of the wafer-scale fabrication process of FTR CMUTs: (i) wafer-scale oxide fusion bonding with circular cavity patterned on the SOI wafer, (ii) wafer thinning through removal of the handling layer, (iii) etching of backside silicon, (iv) membrane patterning, (v) electrode and polyimide patterning, and (vi) Insertion of LMPA and PDMS passivation. b Schematics of the Insertion of LMPA and passivation using PDMS on the backside of a CMUT. The optical images of c outer bent FTR CMUT and d inner bent FTR CMUT (scale bar: 1 cm). e Optical microscopic (OM) image of a CMUT element (scale bar: 500 μm). (Inset) Optical microscopic image of CMUT cells in one element (scale bar: 50 μm). f Cross-sectional scanning electron microscopy image of CMUT composite (scale bar: 1 μm).

Full size image

Electrical characteristics of FTR CMUTs

We confirmed the operation of the FTR CMUTs by measuring the input impedance (Supplementary Fig. 11) using an impedance analyzer and applying both AC and DC voltages on the devices (Fig. 3a). When DC voltage (VDC) was applied from 20 V to 55 V at a step of 5 V, the membrane was deflected by the electrostatic force beyond the static deflection, and the center frequency decreased due to the spring softening effect29,49,50. At VDC of 60 V, pull-in mode was observed where the top membrane deflected beyond the stable state and collapsed down to the bottom insulation layer. The sudden increase in the resonant frequency indicates the pull-in operation mode. At 80% of the pull-in voltage (24 VDC) which is the commonly used operation voltage for CMUTs, the center frequency of 9.6 MHz was observed in air (Fig. 3b). A minor peak around 8.5 MHz was also observed, which is attributed to small variations in membrane radius during fabrication, leading to slightly shifted resonance frequencies across the array. Next, CMUTs were modeled using a 6-component equivalent circuit model, which consisted of a series combination of an inductor (Lx), a capacitor (Cx), and a resistor (Rx) representing the mass, stiffness, and motional resistance of the membrane, respectively. The series branch was connected in parallel with a capacitor (C0) representing the static capacitance of the device. Additionally, to account for parasitic effects from the packaging and electrical interconnections, an extra series branch composed of a resistor (Rs) and a capacitor (Cs) was included (Fig. 3b inset). We also measured the center frequency and pull-in voltage of 60 identical CMUT elements fabricated on the same 4-inch wafer. The variance of center frequency and pull-in voltages was less than 3% and 8%, respectively, which demonstrated the low variability across the fabricated devices (Fig. 3c).

Fig. 3: Electrical characteristics of the FTR CMUTs.
figure 3

a Center frequency in air as a function of DC voltages. b Input impedance of the FTR CMUT at the optimal operational DC voltage (80% of collapsed voltage). (Inset) 6-component modeling of the CMUT and the fitted values of each component. c Center frequencies and pull-in voltages of 60 fabricated FTR CMUTs from the same fabrication run. d Photographs of flexible CMUT undergoing outer and inner bending (scale bar: 3 mm). e Impedance in a flat state and after inner and outer bending fatigue test. f Center frequencies of the device at the flat state and after inner and outer bending fatigue tests.

Full size image

To test the durability of the FTR CMUTs, the bending fatigue tests were conducted using the universal mechanical machine (Fig. 3d). The devices were bent in both inner and outer states with a radius of curvature of 2.5 mm. The fatigue tests consisted of 100 bending cycles with a bending rate of 10 mm/min. As bending was performed in parallel to the 1D rigid structures, no significant changes in the impedance were observed due to the bending (Fig. 3e). The center frequency after the inner and outer bending fatigue tests was within 99% of the initial center frequency at the flat state, which showed the bending durability of the FTR CMUTs (Fig. 3f).

Acoustic characteristics of FTR CMUTs

The acoustic characteristics of the FTR CMUTs including center frequency, fractional bandwidth, transmit acoustic pressure, and receive sensitivity, were measured in oil. First, we applied an input pulse with a peak-to-peak amplitude of 15.8 V and pulse width of 100 ns superimposed on the DC voltage of 48 V to evaluate the impulse response. The measured center frequency was 3.43 MHz in oil with a −3 dB fractional bandwidth of 113%, much larger than that of typical piezoelectric transducers allowing a single CMUT to generate sufficient sound pressure with a wide range of frequencies (Fig. 4a). For stimulation applications, high bandwidth is advantageous since a larger range of frequency can be used to study the effects and the underlying therapeutic mechanism of ultrasound. Next, the output pressure of the FTR CMUTs was measured for different DC voltage and AC voltages (Fig. 4b). At a fixed DC voltage of 48 V, the output pressure linearly increased as expected as the applied input AC voltage increased. A maximum output pressure of 197 kPa was observed in a sinusoidal form with output pressure sensitivities of 12.47 kPa/V, which was measured by a commercial hydrophone at a distance of 20 mm, which is a focal point of the center element of the CMUT array. Similarly, at a fixed AC voltage, the increase in the applied DC voltage also resulted in a higher output pressure because of a higher electrostatic force to actuate the membrane. Moreover, we evaluated the durability of FTR CMUT operation in immersion by measuring the pressure over four consecutive days (Fig. 4c). Similar acoustic pressure was observed during the test without any degradation in acoustic performance.

Fig. 4: Acoustic properties of the FTR CMUTs in immersion.
figure 4

a Impulse response of the CMUT in transmit mode and its frequency response. The hydrophone sensitivity is 197.6 mV/MPa at 3.4 MHz. b Output pressure generated from the CMUT biased at a fixed DC voltage and an increasing AC voltage in transmit mode (black). Output pressure generated from the CMUT biased at a fixed AC voltage and an increasing DC voltage in transmit mode (red). c Continuous operation of the CMUT in transmit mode over four days. d Output voltage of the hydrophone and the CMUT in receive mode, detecting the acoustic waves generated from the commercial transducer. e Output voltage of the CMUT biased at a fixed DC voltage in receive mode in response to various ultrasound pressures generated from the commercial transducer (black). Output voltage of the CMUT biased at various DC voltages in response to acoustic pressure of 70 kPa (red).

Full size image

The receive characteristics of FTR CMUT were also evaluated by measuring the acoustic pressure generated from a commercial transducer, and the CMUT characteristics were compared to those of a commercial hydrophone. For a fair comparison, the FTR CMUT and hydrophone were placed in the same location and distance away from the commercial transducer. At a fixed DC voltage, the measured amplitude of FTR CMUT increased linearly in response to the increasing acoustic pressure generated from the commercial transducer as expected (Fig. 4d, e). The receive sensitivity was estimated to be 332.57 V/kPa at 50 V. Also, we observed the enhancement in the CMUT receive amplitude when biased at higher DC voltages. The performance of both the transmit and receive modes of the CMUT demonstrated robust functionality and reliability, ensuring effective signal transmission and reception even with the FTR structure.

The ultrasound beam generated from the FTR CMUT was also measured and analyzed in the inner and outer bending states. The FTR CMUT was biased at 50 V, and 100 cycles of sine-wave input AC voltage (f = 3.4 MHz, Vpp = 15.8 V) were applied in 1 ms of pulse repetition period. When the FTR CMUT was at the flat state, the maximum ultrasound beam was observed on the surface of each element, which showed similar results to the simulation, showing that the ultrasonic waves generated from the element were not focused on a specific point (Supplementary Fig. 7a, b). However, in the internal bending state, as the radius of curvature decreased, the ultrasound beams were focused at a specific point (Supplementary Fig. 7c (bottom row)). When the 90-degree status was compared to the 30-degree status at the focal point, the axial resolution was five times smaller, and the maximum pressure was 1.6 times higher (Supplementary Fig. 7d). On the other hand, in the outer bending state, the ultrasound beam was dispersed as expected (Supplementary Fig. 7c (top row)). These highly sensitive changes in the ultrasound beam show the importance of controllable curvature in flexible ultrasound transducers.

Dynamic bending modulation of FTR CMUTs

We devised a scheme to modulate the bending of the flexible FTR CMUTs by placing a thin layer of LMPA on the bottom substrate and using direct Joule heating of LMPA to change and fix the curvature of the 1D FTR CMUTs. Since the melting point of Cerrolow 117 is 47.2 °C, at a higher temperature, the curvature of the FTR CMUTs can be modified to a desired bending angle, and once the temperature falls below the melting point, the bending would be fixed. In the FTR structure, the LMPA was placed among the silicon islands and shared the electrical connection with the bottom substrate of the silicon islands, which corresponded to the ground of the CMUT. Thus, by applying DC voltage to both ends of the grounds of the FTR CMUTs, the phase transition in LMPA was induced through Joule heating.

The phase transition was monitored by increasing the applied voltage across the FTR structures (Fig. 5a). At 2.1 VDC, the phase transition from solid to liquid was observed. Moreover, the transient response of Joule heating was monitored by measuring the temperature of the FTR structure with LMPA upon application of a pulse of three different voltages. Immediately after applying a voltage, the temperature quickly increased and saturated at around 90 s (Fig. 5b). At a higher voltage and thus a higher consumed power, the time required to achieve phase transition from solid to liquid decreased as expected (Fig. 5c). We also conducted a cycling test of heating and cooling by applying a pulse of three different voltages. At a higher input voltage, while less time was required to transition from solid to liquid during heating, more time was needed to transition back to solid during cooling. Upon applying 3 V, the device modulation was enabled after an average of 9 seconds and stabilized around 12 seconds after the change (Supplementary Fig. 8). After imposing a specific inner and outer bending to the FTR structures, we confirmed that the bent structures were well maintained without any deformation because of the support from LMPA (Fig. 5d, e). Moreover, the architecture of the FTR CMUT can be adjusted and customized with a high degree of precision and repeatability. The maximum bending angle is approximately 45 degrees. This limit is determined by various factors such as the thickness of the passivation layer and the width of the flexible region between the islands (i.e., element pitch). In addition, we conducted a finite-element-based acoustic simulation of the ultrasound beam generated from the FTR CMUTs bent in inner and outer states with various radii of curvature (width of the device: 12 mm) (Fig. 5f, g). As the radius of curvature of the FTR CMUTs with LMPA in the inner bending state decreased, a higher spatial resolution and a larger acoustic power were observed at the focus.

Fig. 5: Bending modulation of the FTR CMUTs based on the phase transition of LMPA through Joule heating.
figure 5

a Temperature elevation of LMPA at different applied DC voltages via Joule heating. The phase transition temperature of LMPA from solid to liquid is 47.2 °C. b Transient response of the temperature of LMPA when a pulse of three different applied voltages was applied, showing temperature elevation and cooling of LMPA over time. c Consumed power and time required for phase transition at different applied voltages. Photographs of the FTR CMUTs fixed in d outer bending and e inner bending states with various bending radii of curvature (ROC) (scale bar: 1 cm). Simulated ultrasound beam profile of a CMUT array composed of 12 elements placed at a lambda pitch showing normalized pressure generated from the array in f outer bending and g inner bending states.

Full size image

Non-invasive treatment of rheumatoid arthritis using the FTR CMUTs

To demonstrate the potential use of flexible ultrasound (US) transducers for non-invasive treatment of autoinflammatory diseases, we stimulated the spleen of mice using the 1.5-MHz FTR CMUTs on collagen-induced arthritis (CIA) mouse models and showed the therapeutic effects for rheumatoid arthritis (RA) (Supplementary Figs. 9–11). We adopted a long-strip configuration of the flexible CMUT array to facilitate spatially distributed ultrasound stimulation over internal organs, such as spleen. In the widely-used CIA model, autoimmune polyarthritis such as RA is induced through immunization of type II collagen, which is the major constituent protein of articular cartilage51,52,53. To create the CIA model, we immunized mice through injection of an emulsion of complete Freund’s adjuvant and type II collagen (CII) and confirmed the immunization protocols by observing the initiation of clear ankle swelling on all four limbs of CIA mice after 3 weeks of collagen injection (Supplementary Fig. 12). After confirming the model, mice were classified into four groups: RA (-); control cohort, RA (+); model cohort, RA (+), US (+); stimulation cohort, and RA (+), US (−); sham cohort. The 8-week timeline of the experimental procedures consisted of six weeks of disease expression and two weeks of stimulation and evaluation (Fig. 6a).

Fig. 6: Therapeutics effects of ultrasound stimulation of spleen on rheumatoid arthritis (RA).
figure 6

a Timeline of experimental procedures for disease expression and ultrasound stimulation. b Ankle thicknesses before and after ultrasound treatment in 4 cohorts (RA (+), US (+); stimulation groups (RA expression and ultrasound stimulation); n = 10, RA (+) US (−); sham group (RA expression and no stimulation); n = 5, RA (+); model group (RA expression); n = 9, RA (-); control group (wild-type); n = 3). c Thicknesses of four ankles (RF; right forelimb, RH; right hindlimb, LF; left forelimb, LH; left hindlimb) before and after ultrasound treatment in stimulation groups (RA(+), US (+)). d Changes in mice ankle thicknesses over two weeks of treatment in 4 cohorts. *p < 0.05, fisher LSD’s post-hoc test. ns, not statistically significant. e Clinical scores over time after RA expression in 3 cohorts. f Changes in clinical scores over two weeks of treatment in 3 cohorts. g Mice weights over time after RA expression in 3 cohorts.

Full size image

We measured the thickness of four ankles and weights before and after the ultrasound stimulation (i.e., 6-week vs. 8-week time) and compared clinical scores to monitor the therapeutic effects of ultrasound stimulation for rheumatoid arthritis. The ultrasound treatment (f: 1 MHz, power: 350 kPa), which is known to be an effective condition in spleen stimulation38,39, was delivered three times a week over two weeks where each session consisted of 20 minutes of sonication (1 s on and 5 s off). The Intensity (ISPPA; spatial-peak pulse average) used in ultrasound treatment was 4.08 W/cm2 and the Thermal Index (TI) was 5.82. Both parameters were confirmed to be sufficiently safe, within the FDA standards (ISPPA < 190 mW/cm2, TI < 6)54,55. Since our 1.5-MHz transducer offers wide bandwidth (86%), we obtained a sufficiently large intensity at 1 MHz to deliver ultrasound stimulation. We observed reductions in ankle thickness in the CIA mice (n = 10) after the ultrasound stimulation (p = 0.037) (Fig. 6b). Comparing the thickness of the ankle at the 6-week (before US treatment) against those at the 8-week (after US treatment), the overall thickness of the ankles as well as that of each of the four limbs, was reduced. (Fig. 6c). On the other hand, increasing ankle thickness of CIA mice was revealed in the sham cohort. In addition, as expected, no significant changes were observed between the sham and model cohorts (p = 0.62). Considering that the ankle thickness would increase as the mice grow, it was confirmed that the ankle thickness of the control group almost did not change (Fig. 6d, Supplementary Fig. 13). Estimating the clinical score and comparing the score before and after the stimulation, the degree of rheumatoid arthritis was also improved in the stimulation cohort while the symptoms worsened for both the sham and model cohorts (Fig. 6e, f). During the experiments, we monitored the changes in the body weights. Body weights of all three cohorts (stimulation, sham, and model cohorts) continuously increased, indicating the natural growth of mice in all three groups (Fig. 6g).

Gait analysis

We performed gait analysis to monitor any behavioral improvements after ultrasound stimulation. For gait analysis, we created a black tunnel which was placed on a glass plate with green LEDs on its four sides, and a camera was positioned under the glass plate to record the footprints of mice moving in the dark. The LEDs were covered by black insulation tape on the sides to prevent light dispersion. (Fig. 7a–c). The gait performance of the mice on the glass plate was detected clearly and analyzed using DeepLabCut (DLC), which is a versatile and efficient machine-learning (ML)-based toolbox for animal pose estimation (Fig. 7d, e). In general, as the inflammation worsens due to the progression of RA, both the frequency of footprints in touch with the glass plate and the contact area should decrease. To evaluate the effect of ultrasound treatment, we analyzed the frequency and width of footprints in contact with the glass plate and stride length. In the stimulation cohort, the contact frequency increased after the stimulation (Fig. 7f, Supplementary Figs. 14–16), which means that the mice can move a little bit more freely because of the reduction of ankle pain. Similarly, while the footprint width decreased for the sham cohort, no significant changes in the width were observed for the stimulation cohort (Fig. 7g, Supplementary Fig. 17). In addition, the stride length of the stimulation cohort tended to decrease indicating a restoration of gait degradation while no significant changes in the stride length were observed for the sham group (Fig. 7h, Supplementary Figs. 18–20). Notably, it presented a significant increase in the right hindlimb, indicating that the right hindlimb of most mice in the sham group did not contact the glass plate as frequently.

Fig. 7: Gait analysis of RA-induced mice.
figure 7

a Photograph of the experimental setup for gait analysis. (Inset) Photograph of a mouse walking on the glass plate. b Optical image of a footprint while walking. c Image of a footprint with identified toes using DeepLabCut (DLC) program. Identification of each toe (RFC; sole, RF1; index toe, RF2; middle toe, RF3; ring toe, RF4; pinky toe, RH1; big toe, RH2; index toe, RH3; middle toe, RH4; ring toe, RH5; pinky toe) of d right and left forelimb and e hindlimb. Changes of f contact frequency, g footprint width, and h stride length between 6-week and 8-week time for four limbs. *p < 0.05, two-sided unpaired Student’s t-test.

Full size image

Discussion

In this study, we devised a flexible ultrasound transducer by developing a wafer-scale FTR fabrication for CMUTs. FTR structures allowed the flexible transducers to exhibit comparable electrical and acoustic properties to those of standard rigid CMUTs as the transducers were placed in the rigid islands. In addition, we achieved dynamic bending modulation by the use of a low-melting-point metal alloy. Since a small change in the location of elements in an ultrasound array could significantly change the ultrasound beam profile, it is critical to be able to modulate the bending to a known value. Moreover, by using the flexible CMUT with the bending modulation, organ-specific neuromodulation was conducted in vivo for the alleviation of rheumatoid arthritis symptoms. Ultrasound stimulation of the spleen over two weeks resulted in a reduction in the thickness of four ankles and an improvement in the gait. As ultrasound can penetrate deep into our body and can be delivered in a focal manner with controllable beam steering, there is a high potential for wearable ultrasound transducers and therapeutic systems for home healthcare applications. While we demonstrated 1D arrays in this work, future works include the expansion of the FTR fabrication with LMPA bending modulation to 2D ultrasound transducer arrays to enable simultaneous high-quality ultrasound imaging and therapeutics.

In addition, future work will incorporate a more sophisticated and physically realistic COMSOL model to enhance predictive accuracy and design optimization. Despite the modeling limitations, we further evaluated the acoustic output of our CMUT device by comparing it with conventional piezoelectric ultrasound transducers (PUTs) commonly used in neuromodulation. The maximum acoustic pressure of therapeutic PUTs varies widely depending on their intended applications, ranging from approximately 10 kPa to 550 kPa for low-intensity focused ultrasound (LIFU) systems and exceeding 5 MPa for high-intensity focused ultrasound (HIFU) applications41,56,57,58. Within this context, the acoustic pressure generated by our CMUT was within or above the typical LIFU range, suggesting that it is sufficient to induce ultrasound-mediated stimulation for low-intensity therapeutic applications, including spleen stimulation (Supplementary Table 2). While our current device may not yet reach the intensity levels required for HIFU treatment, the acoustic output can be significantly enhanced with further developments, such as optimization of the device geometry, use of advanced membrane materials, and improvement of driving electronics. These advancements would broaden the potential applications of flexible CMUTs, enabling their use not only in neuromodulation but also in a wider range of ultrasound treatments.

Nevertheless, there are still challenges to be addressed in using CMUT in wearable applications. One of the primary disadvantages is the need for precise process control during their manufacturing, coupled with the necessity for high input voltages during operation. These factors can complicate their integration into wearable devices that require low power consumption and ease of use. However, despite these challenges, CMUTs have significant long-term potential for wearable applications. One of the key advantages lies in high compatibility with Complementary Metal-Oxide-Semiconductor (CMOS) technology, which would enable large-scale production and miniaturization. Furthermore, there are various ongoing efforts to reduce the operation voltage of CMUTs, such as by utilizing different membrane materials and pre-charging operation54,55,59. These efforts could potentially address the issue of high voltage requirements for wearable applications.

Methods

Fabrication process of flex-to-rigid structures

To form flex-to-rigid (FTR) structures, the double-sided polishing (DSP) silicon substrate (Unisill Co., Korea) was used for the rigid parts, and LMPA and photo-sensitive polyimide (PI; GPI-102K, PNS technology, Korea) were used for the flexible regions bridging the rigid parts. For the fabrication of silicon islands, thermal oxidation was performed at 1050 °C for 6 min on a DSP silicon wafer using a furnace (E1200, Centrotherm, Blaubeuren, Germany). The backside oxide layer was first patterned using a photoresist and a buffered oxide etch (BOE; Sigma-Aldrich, Saint Louis, USA). Before etching the backside silicon layer, a 50-μm-thick polydimethylsiloxane (PDMS) was coated and cured using a vacuum oven at 85 °C for 4 h on the front side of the Si wafer to protect the front silicon. The backside silicon was anisotropically etched using 5% tetramethylammonium hydroxide (TMAH; Sigma-Aldrich, Saint Louis, USA) with hydroxylamine (NH2OH; Sigma-Aldrich, Saint Louis, USA). The volume ratio of the solution was TMAH:NH2OH = 20:1. After etching the bottom silicon, the PDMS was removed using propylene glycol monomethyl ether acetate (PGMEA; Sigma-Aldrich, Saint Louis, USA) with tetrabutylammonium fluoride solution (TBAF; Sigma-Aldrich, Saint Louis, USA). Next, the wafer was dipped into the BOE solution while covering the backside using a wafer holder to protect the backside SiO2 mask layer (Supplementary Fig. 21). The volume ratio of the solution was PGMEA:TBAF = 99:1. In order to coat polyimide on the frontside wafer, the dummy wafer was bonded to the bottom side using a crystal adhesive bond (821-3 CrystalbondTM 509 Clear, TED PELLA. INC., Redding, USA). A 3-μm-thick photo-sensitive polyimide was coated using a spin coater at 1000 rpm for 20 s and baked at 110 °C for 2 min. The polyimide was exposed using a contact aligner (UV exposure, 200 mJ, 365 nm) and developed in AZ300MIF for 1 min. The patterned polyimide was cured at 250 °C for 2 h. Deep Reactive Ion Etching (DRIE) was performed to remove the remaining thin silicon present in the trench between the rigid islands, and SiO2 acted as the etch stop layer. PDMS was poured on the trench area, and saw dicing was conducted to define each FTR structure.

Fabrication process of flexible FTR CMUTs

For the fabrication of flexible CMUTs in the FTR structures, we started with a highly-doped (boron) double-sided Si wafer ( < 0.005 Ω∙cm) and silicon-on-insulator (SOI) wafer (500-μm-thick device layer (Si), 2-μm-thick box layer (SiO2), and 725-μm-thick handling layer (Si)) (Supplementary Fig. 4). First, 100-nm-thick insulation layers were grown on the Si and SOI wafers through the furnace (wet oxidation at 1050 °C). Next, the vacuum cavities were defined by patterning the thermal oxide layer on the SOI wafer using a 1-μm-thick PR and BOE solution (etch time of 1 min 22 s). To enclose the vacuum cavity with top silicon membranes, the Si wafer and SOI wafer were bonded at the oxide interfaces in a high vacuum using a wafer bonder (SB8 GEN2, SÜSS MicroTec, Garching, Germany). Prior to the wafer bonding, a cleaning step of SPM (H2SO4:H2O2 = 3:1 vol%) at 70 °C was conducted for 10 min followed by SC-1 (H2O:H2O2:NH4OH = 5:1:1 vol%). In order to form a strong bonding at the oxide interface, annealing was conducted in a furnace at 1050 °C in a nitrogen ambient for at least 2 h. The uniform bonding at the interface between oxide layers was confirmed through scanning acoustic microscopy (SAM; Gen6TM C-SAM, Santa Clara, USA). The patterned CMUTs were observed as grey rectangular shapes, and a few defects were observed as black circular and oval shapes (Supplementary Fig. 5). After the wafer bonding, the handling layer was removed by grinding, chemical mechanical process (CMP), and Si wet-etching using TMAH (5%) with NH2OH solution. Next, the silicon islands were created using the fabrication process described in the previous session. To prevent fracturing of the wafer, a 4-inch dummy Si wafer was attached to the bottom using a crystal adhesive bond.

The top silicon was then patterned by the reactive ion etching after patterning with a photoresist to define CMUT membranes. After patterning the membrane, the PR was stripped in the PR stripper (APPS-1) at 50 °C for 10 min. The vias to the ground were formed by etching the top oxide layer using reactive ion etching (RIE). After the elimination of the native oxide layer on the membrane, 10 nm chrome (Cr) and 100 nm gold (Au) layers were deposited using the thermal evaporator to form signal and ground pads. In the same manner as the method of forming the FTR structure, photosensitive polyimide having a thickness of 3 μm was coated and patterned between CMUT elements. After attaching a dummy wafer to the front of the wafer through a crystal bonder to remove the Si layer remaining in the rear, the DRIE process was performed, and SiO2, which had been grown for the insulation layer, was used as the etch stop layer. After the DRIE process, the saw dicing separated the part corresponding to each area of the device with the dummy wafer attached. LMPA was coated on the etched area using a micro syringe needle, at which time the LMPA was placed on a hot plate of 60 degrees or more to be transferred in liquid form. After the LMPA changes to a solid state, soak it in acetone for 30 minutes to detach the dummy wafer by removing the crystal bonder, and clean the IPA and DI for 3 minutes to remove the residue remaining on the surface. To protect against the leakage of LMPA during the bending operation, PDMS was molded to the backside of the CMUT while the front side of the CMUT was protected by water-soluble tape. The PDMS was cured in a vacuum oven at 85 °C for 2 hours, and then to remove water-soluble tape, the device was submerged in the DI for 1 hour. The fabricated FTR CMUTs were connected electrically to the flexible printed circuit board (FPCB) through the wire bonding process (Supplementary Fig. 22). Finally, the CMUTs integrated with the FPCB were dip-coated in a PDMS solution using a dip coater at a coating speed of 0.05 mm/s for encapsulation.

CMUT design

We used the finite element method (FEM)-based COMOSOL simulation to design a single CMUT cell. For analytical calculations, we used the equations shown below to determine resonant frequency and pull-in voltage:

$${f}_{0}=0.469\frac{{t}_{m}}{{R}_{m}^{2}}\sqrt{\frac{\frac{E}{{\rho }_{1}(1-{\sigma }^{2})}}{1+0.67\frac{{R}_{m}{\rho }_{2}}{{\rho }_{1}{t}_{m}}}}\propto \sqrt{\frac{{t}_{m}^{3}}{{R}_{m}^{5}}}$$
$${V}_{{pull}-{in}}=7.542\sqrt{\frac{{g}_{0}^{3}D}{{R}_{m}^{4}{\varepsilon }_{0}}}\propto \frac{{g}_{0}^{\frac{3}{2}}}{{R}_{m}^{2}}$$

where \({R}_{m}\) is the radius of the cavity, \({t}_{m}\) is the thickness of the membrane, \(E\) is Young’s modulus of the membrane, \({\rho }_{1}\) is the density of the membrane, \({\rho }_{2}\) is the density of a surrounding liquid, \(\sigma\) is Poisson’s ratio of the membrane, \({g}_{0}\) is the gap height of the cavity, \({\epsilon }_{0}\) is the vacuum permittivity, and \(D\) is the flexural rigidity of the membrane. The materials used for the membrane, insulation layer, and electrode were Si, SiO2, and Cr/Au, respectively. The designed single CMUT cell consisted of a circular cell with a radius of 11 μm, a 450-nm-thick membrane, and a 100-nm-thick vacuum gap height. The expected pull-in voltage was 60 V, the center frequency was 8.67 MHz in air, and the maximum pressure in water was 175 kPa (Supplementary Fig. 23, Table 1).

Mechanical characterization

The bending tests of the FTR CMUTs were conducted using a universal testing machine (USM-500N, A&D CO., LTD, Japan). The short edges of the FTR CMUT were attached to glass slides which were clamped to the holders in the universal testing machine. The bending speed was 30 mm/min. The FTR CMUT was bent up to the radius of curvature of 2.5 mm at a constant rate. For the fatigue bending test, 100 bending cycles were performed at a constant bending speed in both inner and outer bending with a radius of curvature of 2.5 mm.

Electrical impedance characterization

The electrical properties of CMUTs were analyzed using an impedance analyzer (E4990A, Keysight Technologies, USA) in air. A bias tee consisting of a 1 MΩ resistor and a 22 nF capacitor connected in parallel was used to superimpose the DC voltage on AC voltage during the measurement of impedance in CMUT. By increasing the DC bias voltage in step of 5 V using the high voltage supply (PS310, Stanford Research Systems Inc., Sunnyvale, CA, USA), the operational voltage (80% of collapsed voltage) was confirmed. The components in the equivalent circuit (6-component model) of CMUT were extracted from the impedance analyzer.

Acoustic characterization

To observe the movement of the membrane on elements, the laser Doppler vibrometer (LDV; FV-534, OFV-2570, Polytec GmbH, Germany) was utilized (Supplementary Movie 1). To measure acoustic properties, the FTR CMUTs connected to the FPCB were passivated with PDMS to prevent an electrical shortage of the device in water. The FTR CMUTs were actuated using a function generator (33220A, Agilent Technologies, CA, USA) and an RF amplifier (BT00100-AlphaS-CW, Tomco Technologies, Australia). The generated AC signal was superimposed on a DC signal through a bias tee. The ultrasound generated from the FTR CMUTs was measured using a needle-type hydrophone (NH0500, Precision Acoustics, UK), which was fixed on a programmable motorized stage (Sciencetown Inc., Incheon, Korea). To accurately determine the focal point, three-dimensional scan beam profiling was performed at intervals of 250 micrometers, focusing on a section extending 24 mm from the CMUT surface. The acoustic wave signals were recorded using a high-speed oscilloscope (DSOX2022A, Agilent Technologies Inc., CA, USA).

Bending modulation via joule heating

To confirm the bending modulation based on the phase transition of LMPA via joule heating, a lower voltage than 5 V was applied to the ground pads at both ends of the FTR CMUTs to induce Joule heating. The temperature generated by Joule heating was directly measured by inserting the thermocouple (AT-3K, Hanyoungnux, Korea) into the passivated liquid metal. The phase transition time was confirmed through an optical microscope based on the degree of hardening of the liquid metal.

Collagen emulsion formation

For the formation of a collagen emulsion, we used an immunization-grade chick type II collagen (20012, Chondrex Inc., USA), a Complete Freund’s adjuvant (CFA, 7001, Chondrex Inc., USA), a 3-way stopcock, a 5 ml syringe, a homogenizer (Omni TH Tissue Homogenizer, OMNI International, USA) with a small blade (diameter of 5 mm), and an ice water bath to maintain the emulsion in a cold state during mixing. First, the CFA was added to the syringe, which was sealed with a 3-way stopcock, and then the collagen solution was added gradually at an equal volume of CFA. The emulsion was mixed at 30,000 rpm for 2 minutes before being cooled in an iced water bath for 5 minutes. The mixed collagen emulsion was then dropped into the water to verify that the emulsion was formed correctly (Supplementary Fig. 24).

Collagen-induced arthritis animal model

All animals used in this study were in accordance with protocols outlined and approved by the Institutional Animal Care and Use Committee at the Korea Advanced Institute of Science and Technology (KA2022-022). For the development of CIA, DBA/1 J type (male, 6~7 weeks) mice (n = 31), which are genetically susceptible to autoimmune diseases, were used. The mice were habituated in the animal cage in our lab for a week before the injection of collagen emulsion. Mice were anesthetized with isoflurane (4% induction, 2% maintenance at an O2 flow rate of 1 L/min) using an isoflurane vaporizer system (RWD Life Science Co., Ltd., China), and the 100 μL collagen emulsion was injected subcutaneously into the tail with the tip of the needle inserted between 0.5 and 2 cm from the base of the tail using a 22-gauge Hamilton syringe (1 mL, Gastight Syringe Model 1001 RN, Hamilton®, USA) (Supplementary Movie 2). After 3 weeks from the injection of collagen emulsion, rheumatoid arthritis (RA) was observed through the swelling of ankles. After 6 weeks from the injection, RA was expressed in all of the injected mice (Supplementary Fig. 25a).

Measurement of ankle thickness

To quantify the thickness of swollen ankles due to the expression of rheumatoid arthritis, the thickness of both forelimbs and hindlimbs of 31 mice was measured using a digital thickness gauge (BD547-301, Blue Tec, Korea) every week over 8 weeks. Each ankle was measured three times and the values were averaged. To obtain exact ankle thickness on the same location, the mice were anesthetized with isoflurane (4% induction, 2% maintenance at an O2 flow rate of 1 L/min) on a temperature-controlled warming pad (RT-0515, Kent Scientific Corporation, USA) which was maintained at 37 °C during the measurements (Supplementary Fig. 25b).

Clinical score of rheumatoid arthritis

The thickness of mice ankles was measured for 8 weeks, and each paw was assigned a clinical score for the degree of inflammation from 0 to 4 refers to the scoring system introduced by Chondrex, Inc. (www.chondrex.com/documents/Scoring-System.pdf). For each paw, the clinical score was designated 0 for no erythema and swelling, 1 for mild erythema and swelling that was restricted to the tarsals or ankle joint, 2 for moderate erythema and swelling that extended from the ankle to the metatarsal joints, 3 for severe erythema and swelling that covered the ankle, foot, and digits, or 4 for ankylosis. The maximum scale of clinical scores, adding all paws, was 16 (Supplementary Fig. 25c).

In vivo ultrasound stimulation

For analysis of the effect of the ultrasound stimulation, mice were classified into four groups (stimulation, sham, model, and control cohorts). The sham cohort was subject to the same setup as the stimulation cohort but without ultrasound stimulation. The FTR CMUT integrated with the flexible printed circuit board (FPCB) was positioned on the dorsal skin with ultrasonic gel, slightly lateral to the spinal cord (Supplementary Fig. 26). The ultrasound generated by the FTR CMUTs stimulated the spleen located beneath the dorsal skin (Supplementary Fig. 27). All channels of the CMUT were activated simultaneously by wire-bonding each channel to common signal and ground pads on the FPCB. All channels were biased at a DC voltage of 80 V and were driven by a sine wave at a frequency of 1 MHz with an AC voltage of 17.67 V. The ultrasound with a peak pressure of 350 kPa was delivered for 20 minutes per day (pulse repetition frequency: 0.167 Hz, duty cycle: 16.7%). This treatment was administered 3 times a week for 2 weeks, 6 weeks after the collagen injection. For comparison of the thickness and degree of rheumatoid arthritis, the thickness of all ankles was measured every week.

Gait analysis

A custom-made gait analysis system was used in this experiment. An enclosed walkway consisted of a horizontal glass bottom plate with a blacked-out tunnel on the sides and top, allowing mice to walk through. A green LED light strip was attached to a long edge of the glass bottom plate using Loctite super glue. Black-colored tape was used to cover the opposite long edge of the glass plate to ensure the internal reflection of the light did not escape. A web camera (TC70, TP-Link, China) placed underneath the glass walkway was used to monitor the footprints. Each mouse was placed at the opening of the tunnel to complete three uninterrupted runs along the glass walkway. The experiment was conducted in a dark chamber for conducive mice locomotion. The video was recorded in night view mode with a frame rate of 30 frames/second for behavioral analysis.

DeepLabCut analysis setup

To analyze the gait of the mice in the recorded videos, we utilized DeepLabCut (DLC), a versatile and efficient machine-learning (ML)-based toolbox for animal pose estimation. We employed a GPU-based version of DLC on a Linux system (Ubuntu 16.04), equipped with a 32GB AMD Ryzen 7 1700 8-core CPU and NVIDIA Titan XP GPU. To train our model, we manually labeled each paw and toe for a total of 22 points per mouse as well as 4 reference points from 26 representative videos across 4 batches of experimental data. The ResNet-50 convolutional neural network (CNN) was used to train the system for 800,000 iterations, and the trained DLC network autonomously labeled all the extracted frames from a total of 148 experimental video recordings. The resulting trajectory data was analyzed using a custom-coded MATLAB program to calculate gait characteristics such as the frequency and width of footprints in contact with the glass walkway, and the stride length for each paw and toe (Mathworks®, Natick, MA, USA). The unit for distance was converted into centimeters using the ratio between the actual width of the glass plate and the distance in pixels between the reference points in the video.

Statistical analysis

Statistical analysis was conducted on normally distributed data sets confirmed using Shapiro–Wilk’s normality test. Parametric analysis was performed using a two-sided unpaired Student’s t-test (unequal variances) for behavioral data, and one-way analysis of variance (ANOVA) with Fisher LSD’s post-hoc test for ultrasound treatment data. For one-way ANOVA, the experimental group was used as the factor, and the stimulation, sham, and model groups were used as factors. All data were analyzed using Microsoft Excel (Microsoft) and OriginPro 2019 (OriginLab Co., Northampton, MA, USA).