Augmented ultrasonography with implanted CMOS electronic motes

Modern clinical practice benefits significantly from imaging technologies and much effort is directed toward making this imaging more informative through the addition of contrast agents or reporters. Here, we report the design of a battery-less integrated circuit mote acting as an electronic reporter during medical ultrasound imaging. When implanted within the field-of-view of a brightness-mode (B-mode) ultrasound imager, this mote transmits information from its location through backscattered acoustic energy which is captured within the ultrasound image itself. We prototype and characterize the operation of such motes in vitro and in vivo. Performing with a static power consumption of less than 57 pW, the motes operate at duty cycles for receiving acoustic energy as low as 50 ppm. Motes within the same field-of-view during imaging have demonstrated signal-to-noise ratios of more than 19.1 dB at depths of up to 40 mm in lossy phantom. Physiological information acquired through such motes, which is beyond what is measurable with endogenous ultrasound backscatter and which is biogeographically located within an image, has the potential to provide an augmented ultrasonography.


Input impedance of the piezoelectric transducer
For a piezoelectric transducer, changing the electrical impedance loading the element affects the acoustic impedance. In a similar manner, the mechanical loading affects the electrical impedance. The measured electrical impedance of the 1-mm-by-1-mm-by-0.5-mm lead zirconate titanate (PZT-5A, Piezo Systems) used here is shown in Fig. S1a, where the transducer is mounted on the mote's FR4-based printed circuit board (PCB) submerged in water. The equivalent circuit for this transducer is shown in Fig. S1b. Previous work 4 has shown that with air-backed piezoelectric crystals, a robust inductive band forms between the series and parallel resonant frequencies over a span of approximately 100 kHz for typical element values. However, only a weak resonance is observed here without an inductive band, which is likely due to the lack of proper mechanical boundary conditions that are required for a high-quality-factor mechanical resonance.
Producing such a resonance necessarily complicates the mechanical design of the mote, leading to a higher implant volume. To minimize the overall volume, we instead design the mote to the capacitive impedance response, on the order of 10s of pF for the transducer employed here, which dominates the rest of the impedance spectrum. In this case, the circuit interface to the transducer functions over a much wider frequency range, and smaller transducer sizes can be easily accommodated without changes to the mote circuitry. Fig. S1b shows the capacitive model fit to the impedance data. Figure S1. Input impedance of the piezo element. (a) measured input impedance (real part and imaginary part) of a 1 mm (W) by 1 mm (L) by 0.5 mm (H) lead zirconate titanate mounted on FR4-based PCB submerged in DI water, together with model fit, and (b) the equivalent circuit model producing the fitted curve.

Switch-only rectifier
When the input impedance is dominated by a capacitance on the order of 10s of pF, an inductor value of approximately 100 µH would be required for conjugate matching at a 4-MHz center frequency. Even with bulk ferrites, the smallest such inductor available to date is still in a 0805 package, reducing significantly the ability to scale the total mote volume. To harvest power from the piezoelectric transducers without such inductors, we instead choose to translate one of the techniques widely used in sub-kHz mechanical energy harvesters, the "switch-only" rectifier topology 23 , to the MHz frequency range.
A switch-only rectifier achieves a higher efficiency than simple active rectifiers by increasing the conduction angle when connected to sources whose impedance is dominated by a capacitive component. Fig. S2 illustrates the principle of the switch-only operation, where the transducer is modelled as a capacitor (CS) in series with a current source. With an ideal rectifier, the power from the piezo crystal only flows into the storage capacitor when the voltage across the transducer (VPZ) is equal to VCC, the already stored voltage on the load capacitance, CL. At the end of the conduction phase, the transducer needs to slowly discharge CS, and the next conduction phase begins when VPZ+ is less than -VCC. The switch-only topology adds a switch across the transducer, as well as a pulse generator, shown in blue in Fig. S2. The active diode now is cascaded to the full bridge rectifier, and the output of the comparator from the active diode is used to track the conduction phase. After one conduction phase ends, the switch closes for a short period of time, zeroing the voltage across CS. This effectively shortens the discharge time required to reach the next conduction phase, increases the total conduction time. A theoretical maximum efficiency boost of 100% is predicted 23 , assuming ideal diodes and switches are used to implement such a rectifier. In practice, we simulated an efficiency boost of approximately 46% at the designed load level by the use of this topology compared to optimized passive rectifier-based approaches 12,22 .
Using our measured piezo model and assuming a load resistance of 2.72 GΩ in parallel with a decoupling capacitor, simulation shows the active rectifier can support 1.95 nW of harvested power from a 50-ppm-duty-cycle, 225-kPa-ultrasound pressure source. At this point, the majority of the power is dissipated in voltage clamp (see Fig. S2) within the active rectifier. A load resistance of 180 MΩ corresponds to maximum power transfer; in this case, the active rectifier can harvest up to 8.59 nW.
The 225-kPa-pressure source delivers an incident acoustic power of 835 nW. Acoustic impedance mismatch between soft tissue (~1.54 MRayls) and the piezo crystal (~36 MRayls) reflects 84.3% of this acoustic power, leaving 131 nW for harvesting. Part of this acoustic power is lost in the mechanical to electrical power conversion, given by ! , the piezoelectric efficiency; additional loss comes from the electrical impedance mismatch between the piezo crystal and the rectifier, given by " , the matching efficiency. An air pocket in the package providing backing to the piezo transducer increases ! could be increased by the addition of an air pocket in the package, backing the piezo transducer. Resonant power harvesting can boost " . However, both of these techniques have negative attributes. If an air pocket is included in the mote, its area needs to be the same as that of the piezo, with a thickness beyond one wavelength. In our case, this requires an added volume on the order of 1 mm ´ 1 mm ´ 0.4 mm. The addition of an inductor for resonant power harvesting usually leads to bulky inductors on package. For the values required here, the inductor will be at least be of the scale of 0805 packaging, leading to an extra 2 mm ´ 1.2 mm ´ 1.2 mm in volume. Off-resonance switch-only based power harvesting only requires an on-chip switch, with an area cost in the order of 10s of µm 2 (W/L = 10 µm/0.6 µm transistor used in this work). An off-resonance power harvesting scheme also allows the mote to adopt to a wider range of carrier frequencies.
The interfacing circuits uses thick oxide IO devices for higher voltage compliance. A voltage clamp is added here to clamp VCC to 1.2 V, such that the rectified voltage can be safely used to operate thin oxide (1.8 V core) devices. Figure S2. Simulation results comparing a traditional rectifier and a switch-only rectifier with the switch-only-operation-related circuitry highlighted in blue.

Decision logic for data downlink
The decision logic determines whether the downlink data is a '0' or '1' based on the pulse width received at the mote. Since all pulses in the same pulse packet have the same pulse width, every pulse can be used for downlink recovery. A four-bit count signal indicates the pulse width as the number of half cycles seen at the output of the full bridge rectifier (VFBR) for the pulse that triggers the clock synchronization circuit. The decision logic, in turn, sends this downlink_data as well as an indicator that the downlink data is ready, data_ready, to the higher layers of the protocol stack. A pseudocode description of the decision logic is available as Table S1. To determine downlink_data, the decision logic monitors the historical maximum observed "count" value (value_max). Once count in one frame becomes smaller than value_max by three, value_threshold is then updated for future data decisions, downlink_data is set to 0 if the incoming count value is less than value_threshold, and 1 otherwise, and data_ready is asserted. In addition, if downlink_data has the same value for more than 32 consecutive frames, the decision logic resets itself, such that a new value_max, and thus a new value_threshold can be generated. This "confusion" feature helps to avoid the possibility that value_max is set to an incorrectly high value as a result of bit errors. The link layer protocol uses "0101" as the header to ensure that normal operation will never produce this "confusion" state.

Backscatter modulator
For data uplink, on-off keying of the electrical impedance of the transducer is used to produce a backscatter signal. In Fig. 3f, when the uplink data is 1, we open the NMOS switch MNB, while when the uplink data is 0, MNB is on and shorts the transducer for a maximized data modulation depth. When the transducer is shorted, no power, clock, or down link data can be recovered.
Furthermore, since the generation of the uplink data bit is fully synchronized to the recovered clock, whenever a 0 is sent, the recovered clock is stalled, as MNB remains in the short state, deadlocking the chip. An asynchronous approach is required in the backscatter modulator to prevent this.
The backscatter modulator consists of two matched pulse-generation paths, each consisting of five-bit-tunable, falling-edge-sensitive delay elements as shown in Fig. 3f. Fig. S3 shows one of these pulse generation paths. A falling edge at the input generates an inverted (1-0-1) pulse whose pulse width is set by the delay element. Backscatter modulation logic, the pseudocode of which is given in Table S2, acts as the interface between the synchronous signal uplink_data and the asynchronous pulse generation paths. To ensure robust operation in the case of the transmission of multiple 0s, the logic block generate two signals, data_a (for auxiliary) and data_m (for main).
In the absence of multiple 0s, data_m is the same as uplink_data and data_a stays high.
When a second 0 is to be sent, data_a is pulled low while data_m, at a low level from the previously sent bit 0, resets to 1. With this implementation, each uplink bit 0 takes two clock cycles, or two frames, to transmit. This results from the fact that the bs signal resetting to high asynchronously triggers the clock recovery, which only occurs in the next frame. With such an implementation, transmitting a single bit 0 now takes two clock cycles. To keep a constant data rate in the uplink channel, the backscatter modulator logic artificially delays every bit 1 to also take two clock cycles. As a result, the overall uplink data rate is half of the frame rate.
Another required function of the backscatter modulator logic is to find a working delay value on start-up such that the pulse width for bit-0 transmission is as close to one clock cycle as possible.
Upon power-on reset, signal en is set to 0, which gates off uplink data transmission to allow perfect clock recovery. The backscatter modulator monitors the signal bm and tunes the digital delay value supplied into the delay elements, until a working delay value is found. Finally, signal en is set to 1, enabling uplink data transmission.  Figure S3. The schematic of the five-bit-tunable, falling-edge-sensitive delay elements.

Link layer and application layer
The link layer frame structure is shown in Table S3. INST in the table refers to the four-bit instruction code in the application layer; ARG stands for the eight-bit argument for each instruction code; while RET represents the return value from the mote.

Custom Delay-and-Sum Algorithm
To detect the backscattered data in the reconstructed image, a custom delay-and sum algorithm is necessary to generate floating-point precision images, as the hardware accelerated version For cases in which devices at multiple depths work in parallel, compound reconstruction is performed by first running delay-and-sum multiple times, assuming a focus at the depths of interest (or receive foci), instead of that in the transmit event, and generating multiple B-mode images.
These B-mode images are then combined using four-mm-long rising cosine tapers after envelope detection.

Electrical performance of the chip operating with imaging ultrasound
To fully verify the functionality of the designed integrated circuit with imaging ultrasound, the chip is packaged with the piezoelectric transducer and wirebonded for on-chip logic signal probing.
To verify the chip's performance under realistic ultrasonography situations, B-mode ultrasound pulses are generated using the Verasonics Vantage 256 research ultrasound system with a L12-3V linear array probe (Verasonics Inc.). For the ultrasound imaging session, each frame consists of 192 spatially separate scan lines, or ray lines, with a 100-µs delay between adjacent scan lines. For each scan frame, either a three-cycle-long (0.75 µs) pulse, or a five-cycle-long (1.25 µs) pulse to represent downlink bits 0 and 1, respectively. Fig. S4 shows the hydrophone (Onda Cooperation) measurement of the emitted focused pulse. Nonlinearities in the linear array transducer and its driving waveform are responsible for the "tail" observed in the pulse in Fig. S4. Fig. S5 shows the monitored on-chip supplies during start up when placed within the field of view of the imaging system. The complete start up process takes about 4.7 seconds. When a "QUERY_ID" instruction (with ARG = 0101 0101) is sent from the linear array transducer, the mote correctly recovers the downlink data and delivers the correct uplink frame fully synchronized to the frame rate (Fig. S6). Figure S4. Measured ultrasound waveforms using a hydrophone and an oscilloscope for threecycle, five-cycle, and four-cycle ultrasound pulses, respectively.

Time (s)
which MNB is either on or off. This simulation result is plotted as Figure S7. The equivalent resistance change is between 1.26 GΩ in the off-state and 364 Ω in the on-state. Figure S7. Simulated I-V and derived equivalent electrical resistance of the mote as seen from the piezo crystal with MNB being either on or off.
This loading resistance change on the piezo crystal creates an effective acoustic impedance change, which further induces a backscatter amplitude shift when digital data is sent to the gate of MNB. This change in the amplitude of the backscatter ultrasound can be captured by the Verasonics system. Two recorded scanlines with different uplink data at the same lateral (x-direction) location are compared in Figure S8. Here a difference of 19% in backscatter amplitude at the axial (zdirection) location of the mote is observed. Figure S8. Recorded scanlines at the same lateral (x-direction) location when the uplink data bit is different; the red line showing the case when the uplink data bit is "1" (MNB is open) and the blue line showing the case when the uplink data bit is "0" (MNB is closed).

Image quality degradation in the vicinity of the mote
Since the mote is a strong reflector of ultrasound, images suffer from reduced imaging quality for areas around and below the mote. In the lateral direction, a finite beam width creates an image of the mote larger than its actual width. In the axial direction, the mote blocks the ultrasound energy, dimming the region in the B-mode image below its location. In addition, multiple reflections leads to copies of the mote (reverberant artefacts). The exact amount of interference in the image depends on various parameters, including the beam profile, the dimension of the mote, and the relative position between the mote and the transducer array. For example, side lobes can create horizontal phantom images of the mote if they are strong enough, while a wide main beam convolutes with the mote's dimension, leading to a spread in the width of the mote's image. In the case that the mote is smaller than the width of the beam (either in the axial direction or in the elevational direction), then it does not completely block ultrasound energy, leading to a reduced shadowing of the region below.
To quantify the level of "shadowing" caused by the use of the mote to the image, 3D k-Wave  Figure S9, which shows both reconstructed B-mode images for an intuitive evaluation and the recorded maximum pressure in space as a quantitative evaluation of shadowing.
From the resulting B-mode images, the image of the mote shows minimum lateral spread at the focal depth, while showing up to ±2 mm of spread at 25-mm depth. Reverberant artefacts are no longer present 3 mm below the mote's location. In the pressure recording, the placement of the mote leads to pressure attenuation that shadows the regions below it, as expected; however, this shadowing effect quickly goes away for deeper regions. At approximately 5 mm beneath the mote, the pressure reduction is no longer significant. This shadowing effect is also reduced when the mote is placed in the near field, before constructive interference happens. Figure S9. Phantom setup (2D cross section at zero elevational distance, i.e., y = 0) and 3D ultrasound simulation that estimates the amount of ultrasound energy blocking by the placement of the mote; (a) medium set up with regions of random scatterers; (b) simulation results for reconstructed B-mode image and maximum pressure distribution with no mote (baseline), with the mote placed at (c) 10-mm, (d) 15-mm, (e) 20-mm, and (f) 25-mm distances from the transducer.

Proximity effects from closely placed motes
The operating principle of this augmented ultrasonography allows multiple motes placed within the same field-of-view; however, when they are placed in proximity, data interference may happen that makes them less distinguishable from each other. This is because data from two motes placed within the width of a scanline can get captured and reconstructed into a single point at the resulting B-mode image. Here, a setup shown in Figure S10 is used, with 13 mm thick chicken meat on top of two motes placed with a separation of about 3.3 mm with the imaging transducer array placed 34 mm above the motes. Data signatures captured show spatial separation that is clearly distinguishable with SNRs of 34 dB (left) and 25.9 dB (right). Figure S10. Experiment verifying the ability to distinguish motes places3.3-mm apart at a distance of 34 mm from the source transducer in an in-vitro setup with spatially separate data signatures clearly distinguishable in the reconstructed B-mode image.