A wearable cardiac ultrasound imager – Nature

Materials

Gallium–indium eutectic liquid metal, toluene, ethyl alcohol, acetone and isopropyl alcohol were purchased from Sigma-Aldrich. SEBS (G1645) was obtained from Kraton. Silicone (Ecoflex 00-30) was bought from Smooth-On as the encapsulation material of the device. Silicone (Silbione) was obtained from Elkem Silicones as the ultrasound couplant. Aquasonic ultrasound transmission gel was bought from Parker Laboratories. 1-3 composite (PZT-5H) was purchased from Del Piezo Specialties. Silver epoxy (Von Roll 3022 E-Solder) was obtained from EIS. Anisotropic conductive film cable was purchased from Elform.

Design and fabrication of the wearable imager

We designed the transducer array in an orthogonal geometry, similar to a Mills cross array (Supplementary Fig. 34), to achieve biplane standard views simultaneously. For the transducers, we chose the 1-3 composite for transmitting and receiving ultrasound waves because it possesses superior electromechanical coupling¹⁸. In addition, the acoustic impedance of 1-3 composites is close to that of the skin, maximizing the acoustic energy propagating into human tissues¹⁹. The backing layer dampens the ringing effect, broadens the bandwidth and thus improves the spatial resolution^18,51.

We used an automatic alignment strategy to fabricate the orthogonal array. The existing method of bonding the backing layer to the 1-3 composite was to first dice many small pieces of backing layer and 1-3 composite, and then bond each pair together one by one. A template was needed to align the small pieces. This method was of very low efficiency. In this study, we bond a large piece of backing layer with a large piece of 1-3 composite and then dice them together into small pieces with designed configurations. The diced array is then automatically aligned on adhesive tape with high uniformity and perfect alignment.

Electrodes based on eutectic gallium–indium liquid metal are fabricated to achieve better stretchability and higher fabrication resolution than existing electrodes based on serpentine-shaped copper thin film. Eutectic gallium–indium alloys are typically patterned through approaches such as stencil lithography⁵², masked deposition⁵³, inkjet printing⁵⁴, microcontact printing⁵⁵ or microfluidic channelling⁵⁶. Although these approaches are reliable, they are either limited in patterning resolution or require sophisticated photolithography or printing hardware. The sophisticated hardware makes fabrication complicated and time-consuming, which presents a challenge in the development of compact, skin-conformal wearable electronics.

In this study, we exploited a new technology for patterning. We first screen-printed a thin layer of liquid metal on a substrate. A key consideration before screen printing was how to get the liquid metal to wet the substrate. To solve this problem, we dispersed big liquid metal particles into small microparticles using a tip sonicator (Supplementary Fig. 2). When microparticles contacted air, their outermost layer generated an oxide coating, which lowered the surface tension and prevented those microparticles from aggregating. In addition, we used 1.5 wt.% SEBS as a polymer matrix to disperse the liquid metal particles because SEBS could wet well on the liquid metal surface. We also used SEBS as the substrate. Therefore, the SEBS in the matrix and the substrate could merge and cure together after screen printing, allowing the liquid metal layer to adhere to the substrate efficiently and uniformly. Then we used laser ablation to selectively remove the liquid metal from the substrate to form patterned electrodes.

The large number of piezoelectric transducer elements in the array requires many such electrodes to address each element individually. We designed a four-layered top electrode and a common ground electrode. There are SEBS layers between different layers of liquid metal electrodes as insulation. To expose all electrode layers to connect to transducer elements, we used laser ablation to drill vertical interconnect accesses²¹. Furthermore, we created a stretchable shielding layer using liquid metal and grounded it through a vertical interconnect access, which effectively protected the device from external electromagnetic noises (Supplementary Fig. 8).

Before we attached the electrodes to the transducer array, we spin-coated toluene–ethanol solution (volume ratio 8:2) on the top of the multilayered electrode to soften the liquid-metal-based elastomer, also known as ‘solvent-welding’. The softened SEBS provided a sufficient contact surface, which could help form a relatively strong van der Waals force between the electrodes and the metal on the transducer surface. After bonding the electrodes to the transducer array, we left the device at room temperature to let the solvent evaporate. The final bonding strength of more than 200 kPa is stronger than many commercial adhesives²².

To encapsulate the device, we irrigated the device in a petri dish with uncured silicone elastomer (Ecoflex 00-30, Smooth-On) to fill the gap between the top and bottom electrodes and the kerf among the transducer elements. We then cured the silicone elastomer in an oven for 10 min at 80 °C. As the filling material, it suppresses spurious shear waves from adjacent elements, effectively isolating crosstalk between the elements^18,19. With that being said, we think the main reason for the suppressed spurious shear waves is because of the epoxy in the 1-3 composite, which limits the lateral vibration of the piezoelectric materials. The Ecoflex as the filling material may have contributed but not played a chief role because the kerf is not too wide, only 100 to 200 µm. We lifted off the glass slide on the top electrode and directly covered the top electrode with a shielding layer. Then we lifted off the glass slide on the bottom electrode to release the entire device. Finally, screen-printing an approximately 50-μm layer of silicone adhesive on the device surface completed the entire fabrication.

Characterization of the liquid metal electrode

Existing wearable ultrasound arrays can achieve excellent stretchability by serpentine-shaped metal thin films as electrodes^19,26. The serpentine geometry, however, severely limits the filling ratio of functional components, precluding the development of systems that require a high integration density or a small pitch. In this study, we chose to use liquid metal as the electrode owing to its large intrinsic stretchability, which makes the high-density electrode possible. The patterned liquid metal electrode had a minimum width of about 30 μm with a groove of about 24 μm (Supplementary Fig. 3), an order of magnitude finer than other stretchable electrodes^18,26,57. The liquid metal electrode is ideal for connecting arrays with a small pitch⁵⁸.

This liquid metal electrode exhibited high conductivity, exceptional stretchability and negligible resistance change under tensile strain (Fig. 1b and Supplementary Fig. 4). The initial resistance at 0% strain was 1.74 Ω (corresponding to a conductivity of around 11,800 S m⁻¹), comparable with reported studies^59,60. The resistance gradually increased with strain until the electrode reached the approximately 750% failure strain (Fig. 1b and Supplementary Fig. 4). The relative resistance is a parameter widely used to characterize the change in the resistance of a conductor (that is, the liquid metal electrode in this case) under different strains relative to the initial resistance^58,59,60. The relative resistance is unitless. When the strain was 0%, the initial resistance R₀ was 1.74 Ω. When the electrode was under 750% strain, the electrode was broken and the resistance R at the breaking point was measured to be 44.87 Ω. Therefore, the relative resistance (R/R₀) at the breaking point was 25.79.

To investigate the electrode fatigue, we subjected them to 100% cyclic tensile strain (Fig. 1c). The initial 500 cycles observed a gradual increase in the electrode resistance because the liquid metal, when stretched, could expose more surfaces. These new surfaces were oxidized after contacting with air, leading to the resistance increase (Supplementary Fig. 4). After the initial 500 cycles, the liquid metal electrode exhibited stable resistance because, after a period of cycling, there were not many new surfaces exposed.

This study is the first to use liquid metal-based electrodes to connect ultrasound transducer elements. The bonding strength between them directly decides the robustness and endurance of the device. This is especially critical for the wearable patch, which will be subjected to repeated deformations during use. Therefore, we characterized the bonding strength of the electrode to the transducer element using a lap shear test. The liquid metal electrode was first bonded with the transducer element. The other sides of the electrode and the element were both fixed with stiff supporting layers. The supporting layer serves to be clamped by the tensile grips of the testing machine. Samples will be damaged if they are clamped by the grips directly. Then a uniaxial stretching was applied to the sample at a strain rate of 0.5 s⁻¹. The test was stopped when the electrode was delaminated from the transducer element. A SEBS film was bonded with a transducer element and we performed the lap shear test using the same method. The peak values of the curve were used to represent the lap shear strength (Fig. 1d). The bonding strength between the pure SEBS film and the transducer element was roughly 250 kPa, and that between the electrode and the transducer element was about 236 kPa, which were both stronger than many commercial adhesives (Supplementary Table 2). The results indicate the robust bonding between the electrode and the element, preventing the electrodes from delamination under various deformations. This robust bonding does not have any limitations on the ultrasound pressures that can be transduced.

Characterization of the transducer elements

The electromechanical coupling coefficient of the transducer elements was calculated to be 0.67, on par with that of commercial probes (0.58–0.69)⁶¹. This superior performance was largely owing to the technique for bonding transducer elements and electrodes at room temperature in this study, which protected the piezoelectric material from heat-induced damage and depolarization. The phase angle was >60°, substantially larger than most earlier studies^18,62, indicating that most of the dipoles in the element aligned well after bonding⁶³. The large phase angle also demonstrated the exceptional electromechanical coupling performance of the device. Dielectric loss is critical for evaluating the bonding process because it represents the amount of energy consumed by the transducer element at the bonding interface²⁰. The average dielectric loss of the array was 0.026, on par with that of the reported rigid ultrasound probes (0.02–0.04)^64,65,66, indicating negligible energy consumed by this bonding approach (Supplementary Fig. 1b). The response echo was characterized in time and frequency domains (Supplementary Fig. 1c), from which the approximately 35 dB signal-to-noise ratio and roughly 55% bandwidth were derived. The crosstalk values between a pair of adjacent elements and a pair of second nearest neighbours have been characterized (Supplementary Fig. 1d). The average crosstalk was below the standard −30 dB in the field, indicating low mutual interference between elements.

Characterization of the wearable imager

We characterized the wearable imager using a commercial multipurpose phantom with many reflectors of different forms, layouts and acoustic impedances at various locations (CIRS ATS 539, CIRS Inc.) (Supplementary Fig. 11). The collected data are presented in Extended Data Table 1. For most of the tests, the device was first attached to the phantom surface and rotated to ensure the best imaging plane. Raw image data were saved to guarantee minimum information loss caused by the double-to-int8 conversion. Then the raw image data were processed using the ‘scanConversion’ function provided in the k-Wave toolbox to restore the sector-shaped imaging window (restored data). We applied five times upsampling in both vertical and lateral directions. The upsampled data were finally converted to the dB unit using:

The penetration depth was tested with a group of lines of higher acoustic impedance than the surrounding background distributed at different depths in the phantom. The penetration depth is defined as the depth of the deepest line that is differentiable from the background (6 dB higher in pixel value). Because the deepest line available in this study was at a depth of 16 cm and was still recognizable from the background, the penetration depth was determined as >16 cm.

The accuracy is defined as the precision of the measured distance. The accuracy was tested with the vertical and lateral groups of line phantoms. The physical distance between the two nearest pixels in the vertical and lateral directions was calculated as:

We acquired the measured distance between two lines (shown as two bright spots in the image) by counting the number of pixels between the two spots and multiplying them by Δy or Δx, depending on the measurement direction. The measured distances at different depths were compared with the ground truth described in the data sheet. Then the accuracy can be calculated by:

The lateral accuracy was presented as the mean accuracy of the four neighbouring pairs of lateral lines at a depth of 50 mm in the phantom.

The spatial resolutions were tested using the lateral and vertical groups of wires. For the resolutions at different depths, the full width at half maximum of the point spread function in the vertical or lateral directions for each wire was calculated. The vertical and lateral resolutions could then be derived by multiplying the number of pixels within the full width at half maximum by Δy or Δx, depending on the measurement direction. The elevational resolutions were tested by rotating the imager to form a 45° angle between the imager aperture and the lines. Then the bright spot in the B-mode images would reveal scatters out of the imaging plane. The same process as calculating the lateral resolutions was applied to obtain the elevational resolutions. The spatial resolutions at different imaging areas were also characterized with the lateral group of wires. Nine wires were located at ±4 cm, ±3 cm, ±2 cm, ±1 cm and 0 cm from the centre. The lateral and axial resolutions of the B-mode images from those wires were calculated with the same method.

Note that the lateral resolution worsens with the depth, mainly because of the receive beamforming (Supplementary Fig. 15). There are two beamformed signals, A and B. The lateral resolution of the A point (x₁) is obviously better than that of the B point (x₂). The fact that lateral resolution becomes worse with depth is inevitable in all ultrasound imaging, as long as receive beamforming is used.

As for different transmit beamforming methods, the wide-beam compounding is the best because it can achieve a synthetic focusing effect in the entire insonation area. The better the focusing effect, the higher the lateral resolution, which is why the lateral resolution of the wide-beam compounding is better than the other two transmit methods at the same depth. Furthermore, the multiple-angle scan used in the wide-beam compounding can enhance the resolution at high-angle areas. The multiple-angle scan combines transmissions at different angles to achieve a global high signal-to-noise ratio, resulting in improved resolutions.

The elevational resolution can only be characterized when the imaging target is directly beneath the transducer. For those targets that are far away from the centre, they are difficult to be imaged, which makes their elevational resolutions challenging to calculate. When characterizing the elevational resolution, the device should rotate 45°. In this case, most of the reflected ultrasound waves from those wires cannot return to the device owing to the large incidence angles. Therefore, those wires cannot be captured in the B-mode images. One potential solution is to decrease the rotating angle of the device, which may help capture more wires distributed laterally in the B-mode image. However, a small rotating angle will cause the elevational image to merge with the lateral image, which increases the error of calculating the elevational resolution. Considering those reasons, we only characterized the elevational resolution of the imaging targets directly beneath the transducer array.

The contrast resolution, the minimum contrast that can be differentiated by the imaging system, was tested with greyscale objects. The collected B-mode images are shown in Fig. 2. Because the targets with +3 and −3 dB, the lowest contrast available in this study, could still be recognized in the images, the contrast resolution of the wearable imager is determined as <3 dB.

The dynamic range in an ultrasound system refers to the contrast range that can be displayed on the monitor. The contrast between an object and the background is indicated by the average grey value of all pixels in the object in the display. The grey value is linearly proportional to the contrast. The larger the contrast, the larger the grey value. Because the display window was using the data type ‘uint8’ to differentiate the greyscale, the dynamic range was defined as the contrast range with a grey value ranging from 0 to 255.

The object with −15 dB contrast has the lowest average grey value, whereas the object with +15 dB contrast has the highest (Supplementary Fig. 16). In our case, there are six objects with different contrasts to the background in the phantom. The highest grey value obtained from the object of +15 dB contrast was 159.8, whereas the lowest grey value from the object of −15 dB contrast was 38.7. We used a linear fit to extrapolate the contrasts when the corresponding average grey values were equal to 255 and 0, which corresponded to contrasts of 39.2 dB and −24.0 dB, respectively. Then the dynamic range was determined as:

The dead zone is defined as the depth of the first line phantom that is not overwhelmed by the initial pulses. The dead zone was tested by imaging a specific set of wire phantoms with different depths right beneath the device (Supplementary Fig. 11, position 4) directly and measuring the line phantoms that were visible in the B-mode image.

The bandwidth of the imager is defined as the ratio between the full width at half maximum in the frequency spectrum and the centre frequency. It was measured by a pulse-echo test. A piece of glass was placed 4 cm away from the device and the reflection waveform was collected with a single transducer. The collected reflection waveform was converted to the frequency spectrum by a fast Fourier transform. The full width at half maximum was read from the frequency spectrum. We obtained the bandwidth using:

Contrast sensitivity represents the capability of the device to differentiate objects with different brightness contrasts²⁰. The contrast sensitivity was tested with the greyscale objects. The contrast sensitivity is defined as the contrast-to-noise ratio (CNR) of the objects having certain contrasts to the background in the B-mode image:

in which μ_in and σ_in are the mean and the standard deviation of pixel intensity within the object, and μ_out and σ_out are the mean and the standard deviation of pixel intensity of the background.

The insertion loss is defined as the energy loss during the transmission and receiving. It was tested in water with a quartz crystal, a function generator with an output impedance of 50 Ω and an oscilloscope (Rigol DS1104). First, the transducer received an excitation in the form of a tone burst of a 3-MHz sine wave from the function generator. Then the same transducer received the echo from the quartz crystal. Given the 1.9-dB energy loss of the transmission into the quartz crystal and the 2.2 × 10⁻⁴ dB (mm MHz)⁻¹ attenuation of water, the insertion loss could be calculated as:

Simulation of the acoustic field

The simulation computes the root mean square of the acoustic pressure at each point in the defined simulation field. The root mean square is defined in the equation below and gives an average acoustic pressure over a certain time duration, which is pre-defined in a packaged function of the software. In the equation, x_i is the simulated acoustic pressure at the ith time step.

Figure 2c is the simulated root mean square of the transmitted acoustic pressure field by the orthogonal transducers. The simulation was done using the MATLAB UltraSound Toolbox⁶⁷. Each one-dimensional phased array in the orthogonal transducers gives a sector-shaped acoustic pressure field. The simulation merges two such sector-shaped acoustic pressure fields. The imaging procedure was done with the same parameters as the simulations.

In the simulation, we defined the transducer parameters first: the centre frequency of the transducers as 3 MHz, the width of the transducers as 0.3 mm, the length of the transducers as 2.3 mm, the pitch of the array as 0.4 mm, the number of elements as 32 and the bandwidth of the transducers as 55%. Then we defined wide-beam compounding (Supplementary Fig. 13) as the transmission method: 97 transmission angles, from −37.5° to +37.5°, with a step size of 0.78°. Then the acoustic pressure field was the overall effect of the 97 transmissions. Finally, we defined the computation area: −8 mm to +8 mm in the lateral direction, −6 mm to +6 mm in the elevational direction and 0 mm to 140 mm in the axial direction.