Main

Linking brain function across scales—from cells to circuits and eventually to systems—is a major challenge. Functional magnetic resonance imaging (fMRI) indirectly maps neuronal activity across the entire brain, based on vascular hemodynamics, for example, blood flow, blood volume or blood oxygenation levels1,2,3,4. To characterize the neuronal basis of fMRI signals, the MR-compatible electroencephalogram (EEG) preamplifier has been implemented inside the MRI scanner during scanning5, setting an ingenious scheme for concurrent EEG–fMRI acquisition. Simultaneous EEG–fMRI has been increasingly used to monitor both neuronal and hemodynamic activities6,7, explaining the neurovascular coupling mechanisms underlying the hemodynamic response detected by fMRI in healthy and diseased brains8,9. EEG offers temporal information about brain state dynamics while fMRI maps the spatial distribution of brain state network, making their combination powerful for localizing disease foci10 and tracking brain networks during resting state11, sleep12 or cognition13.

Despite steady progress over the two decades, simultaneous EEG–fMRI remains technically demanding. Recording tiny EEG signals inside an MR scanner requires reliable cable connections to create common voltage ground between the skull, EEG receiver and scanner bore9. This multi-cable configuration requires careful installation, because even slight changes in junction resistance may alter the baseline, making it challenging to reproducibly retrieve a baseline pattern that is necessary for follow-up analysis. Also, the wired connections between recording electrodes and EEG receiver collect electromagnetic interference artifacts14, especially during the MR excitation pulses and switching magnetic field gradients15. While artifacts can partially be removed during postprocessing16,17,18,19, reliable recovery of weak EEG signals from the much stronger background interference requires both high-gain preamplifiers and high-speed analog digital converters with large dynamic ranges20, necessitating extra shielding and powering modules, increasing system complexity and safety concerns21,22,23.

Another approach reduces artifacts by synchronizing the EEG apparatus with the MR scanner thus acquiring EEG signals only during the MRI acquisition window with switched-off excitation pulse and stable encoding gradient24, but this requires precise synchronization and previous knowledge of the pulse sequence, which can be difficult to implement. Wireless electrophysiology transducers25,26 have been developed to use on-board gradient sensors and microcontrollers for dynamically identifying proper acquisition windows with stable gradients, enabling the synchronously acquired EEG signals to be encoded onto the wireless carrier wave and detected by a standard MRI coil. While promising, this approach requires on-board microcontrollers and dedicated radio frequency (RF) transmitters powered by sizable internal batteries, limiting miniaturization for interventional applications. In addition, this transducer could only record EEG signals when the gradients are stable, leading to a larger fractionally reduced effective acquisition time. All these challenges preclude widespread EEG–fMRI usage.

To overcome these limitations, we developed a wirelessly powered oscillator encoding fMRI and EEG signals simultaneously. This is an advance over wireless amplified NMR detectors for deep tissue imaging27 that combined wirelessly powered amplifiers28 with miniaturized micro-coils29,30. Here, when the wireless pumping power exceeds an oscillation threshold28, the wireless amplified NMR detectors operates as an oscillator that can directly convert wirelessly provided pumping power into sustained oscillation currents near the circuit’s resonance frequencies. Unlike conventional voltage-controlled oscillators that can only encode low-frequency signals after down-conversion31, the wireless oscillator leverages circuit nonlinearity to combine down-conversion and frequency encoding of MRI signals into a single stage32. Because circuit oscillation can also be modulated by low-frequency bias voltages applied on its nonlinear components33, low-frequency neuronal signals are also encoded onto the same frequency-modulated carrier wave, but on a distinct sideband from the simultaneously encoded high-frequency MRI signals.

The oscillation carrier wave can be wirelessly detected by a standard MRI coil throughout MR acquisition, in the same manner as how conventional MR signals are detected. Without the need for dedicated gradient sensors or synchronization apparatus, the oscillator can reliably encode MRI and EEG signals, even during gradient switching periods. Since the down-converted MRI signals and neuronal signals exhibit different frequency separations from the carrier center, they can be told apart by high-pass and low-pass filtering after frequency demodulation, eliminating the need for synchronization hardware. The pumping power can reduce the circuit’s effective resistance and increase its quality factor by ~39,000-fold, making the circuit’s oscillation frequency sensitive to small modulation voltages, thus obviating the need for high-power preamplifiers or digitizers that were traditionally required to recover subtle neuronal signals from the artifactual background. Without the need for analog digital converters or microprocessors, the compact oscillator implants easily on the skull, requiring only a few milliwatts of wireless power to activate the transducer, inducing negligible heating effects.

In this study, we first fabricated a wireless integrated sensing detector for simultaneous EEG and MRI (WISDEM) capable of retrieving low and high-frequency signals from frequency-modulated-encoded sidebands. We then tested the feasibility and performance of WISDEM to retrieve low-frequency voltage signals when a train of sinusoidal waves was directly injected into the sensing electrodes. Furthermore, we validated the detector’s imaging performance by observing robust echo planar imaging-blood-oxygenation-level dependent (EPI-BOLD) performance in the S1 forepaw region (S1FP) in rodents on electrical forepaw stimulation. Last, we combined optogenetic stimulation with simultaneous acquisition of local field potential (LFP) and fMRI signals to expand the applicability of WISDEM. These results demonstrate the utility of this hybrid two-in-one detector for simultaneous EEG–fMRI signal recordings of rodent brains through the MR console system, paving the way for studying cross-bandwidth neuronal and hemodynamic responses.

Results

Operation principle and circuit fabrication

The parametric resonator (PR) is a circular-shaped loop-gap resonator with a continuous center conductor to bridge its virtual grounds. As a result, the resonator has a circular resonance (cr) mode (Fig. 1a) at a lower frequency ωcr and a butterfly resonance (br) mode (Fig. 1b) at a higher frequency ωbr. When a pumping signal is applied at approximately the sum frequency of these two modes ωbr + ωcr, the resonator can oscillate at frequencies (ωb and ωc) that are close to the resonance frequencies (ωbr and ωcr) of individual modes, that is, ωb ≈ ωbr and ωc ≈ ωcr. Once the pumping signal ωp is determined by an external frequency synthesizer, it will also determine the sum of butterfly and circular oscillation frequencies, that is, ωp = ωb + ωc. If the butterfly-mode oscillation signal falls within the detection band of the MRI scanner, it can be detected by a standard MRI coil. As explained in the Methods, the oscillation frequency ωb has a linear relation with the resonance frequency ωbr. If at the same time, the butterfly mode also interacts with an MRI signal that is separated from ωb by an offset Δf that is smaller than the imaging bandwidth, the MRI signal will interact with the oscillation signal, creating a down-converted signal at Δf that can frequency modulate the oscillation signal at the same time32. In this way, both the low-frequency EEG signal and the high-frequency MRI signal can be encoded onto the same carrier wave for wireless transmission.

Fig. 1: The WISDEM detector consisted of a PR overlaid by a VSR.
figure 1

a, The PR had a circular-shaped conductor pattern (orange) that was split by a pair of varactor diodes (purple) connected in head-to-head configuration, creating a resonance mode with the circular-shaped current flow. b, The PR also had a continuous center conductor to create a second resonance mode with the butterfly-shaped current flow. c, The VSR was made by consecutively wrapping an enameled copper wire (pink) around two parallel rods, thus creating five counterclockwise turns in the first rod and another five clockwise turns in the second rod. The wire’s two end edges were soldered to the emitter and collector terminals of a BJT (labeled by N terminals in the green rectangle). A pair of electrodes (yellow) for sensing and grounding were connected to the transistor’s base and emitter through 10-kΩ resistors (red). d, The VSR was overlapping across the edge of the PR, with one loop sitting inside the conductor pattern of the PR and the other loop sitting outside, thus creating effective coupling with only the circular-mode resonance of the PR. e,f, When a bias voltage was applied across the pair of electrodes (e), it linearly shifted the oscillation frequency at a rate of 5.5 kHz mV−1 (f). This FVR was 55-fold larger than the 3-dB linewidth of the oscillation peak (~100 Hz as shown in the inset).

Source data

Full size image

The PR had a resonance mode at 399.5 MHz (Q = 79) with circular-shaped current flow (Fig. 1a). By connecting the two virtual voltage grounds of the circular mode with a horizontal conductor, a second resonance mode was created at ωbr = 300.2 MHz (Q = 77) with butterfly-shaped current flow (Fig. 1b). Because the horizontal conductor was connecting the virtual voltage grounds of the circular mode, introduction of the second resonance mode will hardly affect the first resonance mode. The voltage sensing resonator (VSR) had a figure of eight conductor pattern (Fig. 1c) that was connected to the emitter and collector terminals of a bipolar junction transistor (BJT). The BJT’s base was connected to its emitter via a 475-kΩ resistor that can neutralize excessive charge accumulated on the BJT’s base while maintaining sufficient internal impedance for the transducer. By connecting the BJT’s base with a sensing electrode via a 10-kΩ resistor and the BJT’s emitter with a grounding electrode via another 10-kΩ resistor, the resonance frequency of the VSR can be effectively modulated by the bias voltage applied across the electrode pair. Meanwhile, the two 10-kΩ resistors (Fig. 1c) can effectively isolate the entire RF circuit from the sensing electrodes that directly touch biological tissues, thus improving circuit stability. According to the voltage division relation, these two 10-kΩ resistors will only reduce the sensing voltage by a factor of 4% when they are serially connected to the internal impedance of the transducer that is mostly defined by the 475-kΩ resistor between the base and the emitter.

When the PR was enclosed inside an enhancer resonator that could concentrate pumping magnetic field, the PR’s circular-mode resonance frequency decreased to ωcr = 380.8 MHz (Q = 79) while its butterfly-mode resonance frequency remained unchanged at 300.2 MHz. Concurrently, the enhancer’s resonance frequency was adjusted to 676 MHz, which was slightly below the sum of butterfly-mode resonance frequency (ωbr = 300.2 MHz) and circular-mode resonance frequency (ωcr = 380.8 MHz). Meanwhile, when the VSR was overlapping across the circular edge of the PR (Fig. 1d,e), with one loop sitting inside the PR and another loop sitting outside, the VSR could couple with the circular mode of the PR and decreased the circular-mode resonance frequency to 374.8 MHz (Q = 67). It is worthwhile mentioning that both circles of the VSR were symmetrically distributed across the center conductor line of the PR, and the VSR’s interaction with the butterfly mode of PR was effectively canceled. As a consequence, the VSR was interacting with only the circular mode of the PR, enabling effective modulation of the oscillation frequency. When a pumping signal was provided at 675.0 MHz by a loop antenna, the PR had sustained oscillation current at ωb = 300.2 MHz and ωc = 374.8 MHz. When we varied the d.c. bias voltage across the pair of sensing electrodes, the oscillation signal shifted at a rate of 5.5 kHz mV−1 (Fig. 1f). This rate of frequency shift was defined as the frequency-to-voltage ratio (FVR). The narrow linewidth (~100 Hz, Fig. 1f) of oscillation peak compared to large voltage-induce frequency shift would enable the wireless detector to identify input voltages as small as 18 µV.

Retrieving low-frequency input voltages

To simulate neuronal input signals, we directly injected waveforms produced by a function generator into the sensing electrodes. The function generator produced 20 pulses every other second. Each pulse had a duration of 20 ms, corresponding to one complete sinusoidal cycle. A 10-mm Bruker surface coil was placed behind the WISDEM to relay the oscillation signal into the scanner console (Fig. 2a).

Fig. 2: Retrieving low-frequency voltage signals applied on the sensing electrodes.
figure 2

a, The schematic setup to characterize the frequency response of the WISDEM detector when a train of sinusoidal waves were directly injected into the sensing electrode and the ground reference. b, The sinusoidal waveform was reconstructed by derivatizing the phase of oscillation signal and dividing this phase derivative with the FVR, that is, \(({{\mathrm{d}}\varnothing }_{t}/{{\mathrm{d}}t})\)/FVR. c, Zoom-in view of five epochs (20 sinusoidal pulses each). Stacked waveforms (light gray, n = 5) closely matched the simulated input (blue), with the averaged profile (red) showing in high agreement (insert at the bottom left of the figure). d, Top: bar plots showing the distribution of reconstructed peak intensities for each applied voltage (mean ± s.d., n = 100 waveforms). Bottom: box and whisker plots summarizing the same distribution of peak intensities: where whiskers denote the minimum and maximum, and boxes show the mean (center line) and interquartile range.

Source data

Full size image

Once the complex oscillation signal was recorded by the quadrature receiver chain of the spectrometer, its instantaneous phase angle was obtained by the angle and unwrap functions in MATLAB. The frequency shift at each time point was obtained from the phase derivative of the oscillation signal, by multiplying the spectrometer sampling speed (326,087 Hz) with the phase difference between adjacent sampling points. After low-pass filtering the frequency shift below a 1-kHz passband, the input waveform (Fig. 2b) was obtained by dividing this frequency shift with the oscillator’s FVR (5.5 kHz mV−1). The retrieved waveform (red in Fig. 2c) was averaged over all five epochs, which agreed well with the input waveform. When we varied the input waveform intensity, we observed a 1:1 linear relation (Fig. 2d) between the peak input voltage and the peak voltage value of the reconstructed waveform. This high-level consistency demonstrates reliable voltage encoding capability of the WISDEM.

Retrieving high-frequency MR signals

To evaluate the detector’s performance for MR signal encoding using an EPI sequence, we placed the detector on an agarose phantom (1% agarose dissolved in distilled water). We then adjusted the pumping power to 0.4 dBm above the oscillation threshold28 and continuously acquired the oscillation signal during the entire EPI acquisition period. The horizontal field of view (FOVx = 26 mm) was enlarged to threefold the vertical FOVy (78 mm) so that the spectral window was large enough to include the information-encoding sidebands, achieving object recovery in the absence of quadrature frequency-modulated encoding. By empirically adjusting the pumping frequency, the oscillation signal was adjusted to ~81.5 kHz above water resonance and aligned to the left quarter location in the frequency domain, thus separating it from the image center by one-quarter of the horizontal FOVx (Fig. 3a). In this way, the MR signal had the largest distance separation from its mirror and the aliased mirror. Compared to neuronal voltages that modulated the oscillation signal at <1 kHz speed, MRI signals modulated the oscillation signal at the offset frequency (for example, ~81.5 kHz), showing up as distinct sidebands in the frequency domain.

Fig. 3: Retrieving high-frequency MR signals.
figure 3

a, During the gradient encoding periods of EPI sequence, low-frequency EEG signals and high-frequency MRI signals were simultaneously encoded onto the same oscillation carrier wave that a standard MRI coil could directly detect. The oscillation carrier frequency was adjusted to overlap with the left quarter location (represented by the white dash line) of the FOV in the frequency domain, so that the MR signal was separated from its mirror and the mirror’s aliasing by largest distances. The phase derivative value was then low-pass filtered to obtain EEG and high-pass filtered to obtain MRI. b, A phantom image reconstructed from the oscillation signal recorded during the EPI sequence. c, The SNR of WISDEM (up to 135) in its sensitivity profile along the orange dashed line crossing through the image center in b. d, The phantom image was obtained by a reference coil with direct cable connection to the scanner console. e, The SNR of the reference coil (up to 222) in the sensitivity profile along the orange dashed line crossing through the image center in d. FFT, fast Fourier transform; HPF, high-pass filtered; RF, radio frequency.

Source data

Full size image

We derivatized the phase \({\varnothing }_{t}\) of the oscillation signal at each time point and assigned the high-pass filtered (HPF) value of this phase derivative \({{\mathrm{d}}\varnothing }_{t}/{{\mathrm{d}}t}\) as the amplitude signal At for that particular time point, that is, \({A}_{t}={\mathrm{HPF}}({{\mathrm{d}}\varnothing }_{t}/{{\mathrm{d}}t})\) (Fig. 3a). Meanwhile, because the position of the oscillation sideband for MR signal is determined by Larmor frequency, the oscillation signal will accumulate a phase angle \({\varnothing }_{t}\) that is equal to the sideband’s frequency separation integrated over the evolution time, where \({\varnothing }_{t}\) can be directly measured by the spectrometer. During the imaging sequence, a complex signal for a particular time point can be obtained by multiplying the signal amplitude At with the phase term \(\exp ({-j\varnothing }_{t})\), leading to \(\exp ({-j\varnothing }_{t})\times {\mathrm{HPF}}({{\mathrm{d}}\varnothing }_{t}/{{\mathrm{d}}t})\). After applying two-dimensional (2D) Fourier transformation on phase-sensitive signal series, the image slice were retrieved (Fig. 3b). Because the mirrored signal and aliased mirror had opposite phase relations with respect to the MR signal, they appeared as a dispersed pattern after 2D Fourier transformation and were not used in subsequent analysis.

To evaluate image sensitivity, we repeated the same procedure to obtain a second image (S2) and calculated signal-to-noise ratio (SNR) of individual pixels by dividing the average intensity of individual pixels with the standard deviation of background signal intensity in the difference image.

$${\mathrm{SNR}}=\frac{\left(|{S}_{1}|+|{S}_{2}|\right)/2}{{\mathrm{s.d.}}\left(|{S}_{1}|-|{S}_{2}|\right)}$$

The image (Fig. 3b) reconstructed from the oscillator had maximum SNR of 135 (Fig. 3c) along the orange dashed line passing through the image center. For comparison, we also constructed a reference coil of the same dimension (Extended Data Fig. 1) but with a direct wired connection to the scanner console. The reference image (Fig. 3d) obtained by this cable-connected coil has a similar shape to Fig. 3b, but a maximum SNR of up to 222 (Fig. 3e) crossing through the image center. When comparing their maximum SNR through the image center, the WISDEM maintains about 60% sensitivity of a cable-connected coil. This SNR reduction comes primarily from the PR whose frequency-modulated-decoded signal amplitude needs to be multiplied by the phase for correct image reconstruction in the center portion of the extended FOV. The center cropped regions were used for subsequent image analysis.

Retrieving BOLD signals during forepaw stimulation

Next, we demonstrated the detector’s capability of recording BOLD signals in vivo by activating rat’s somatosensory cortex via stimulating its left forepaw inside an MRI scanner (Fig. 4a) with 0.33-ms biphasic electric pulses, 2 mA, 5 Hz for 4 s, followed by an 11-s resting period to restore the activated brain region to its resting state baseline (Fig. 4b). Concurrently, during forepaw stimulation, we activated the WISDEM by a pumping signal at 1 dBm beneath the detector’s oscillation threshold so that we could repetitively acquire EPI with enhanced amplitude and high sensitivity27. After the amplified images were aligned with the standard SIGMA brain atlas34, we calculated the correlation coefficient between the time-dependent signal intensity of each pixel and the ideal stimulation function, thus identifying brain regions that were activated by forepaw stimulation. As shown in the color-coded activation maps (Fig. 4c) overlapped on the atlas34, the S1FP region had a BOLD modulation pattern with the same periodicity (15 s) as the stimulation epoch, which was consistent with previous results35,36,37 using conventional surface coils. Within each epoch, the MR signal intensity had up to 1.5% modulation during the 4-s stimulation period before returning to the baseline (Fig. 4d).

Fig. 4: EPI-based fMRI using the WISDEM from rats upon electrical forepaw stimulation.
figure 4

a, Image of the WISDEM detector placed on top of a rat head that was secured in a cradle. b, Schematic of the rat inside the MRI scanner receiving electrical stimulation to the left forepaw (8 epochs of 333-µs biphasic pulses at 5 Hz for 4 s, followed by 11 s of rest). c, BOLD fMRI activation maps show responses in S1FP region on stimulation, aligned to the SIGMA rat brain atlas34 (n = 5 animals, single-sided GLM-based t-statistics in AFNI, P(FDRcorrected) = 0.0001, where FDR means false discovery rate). d, S1FP signal modulation synchronized with stimulation epochs (pink dashed line). Single-sided GLM-based t-statistics in AFNI were used. P(FDRcorrected) = 0.0001.

Source data

Full size image

Simultaneously retrieving BOLD and LFP signals

To demonstrate the full-scope capability of the WISDEM, we simultaneously recorded LFP–MRI. Rats were stimulated by light pulses (470-nm wavelength) with an optical fiber and an electrode inserted into the S1FP region transfected with AAV5-CaMKII.hChR2 (Fig. 5a,b). To confirm LFP recording, we acquired the detector’s oscillation signals without encoding gradients or RF pulses during 2 Hz of light stimulation (3 pulses) with a 1-s rest interval. Then, we derivatized the phase angle of the complex oscillation signal and divided the phase derivative \({{\mathrm{d}}\varnothing }_{t}/{{\mathrm{d}}t}\) with the FVR (5.5 kHz mV−1) to obtain time-dependent electrophysiological signals. EEG spikes were observable for low-power light at 0.39 mW (Fig. 5c). In a control experiment, reducing light power to 0.005 mW eliminated LFP patterns (Fig. 5c), linking negative peaks to light induced activity.

Fig. 5: WISDEM for optogenetic-fMRI in rats.
figure 5

a, The representative histology of the brain slice and fluorescence images showed AAV-ChR2 expression in S1FP region (n = 4 animals). Scale bars: 1 mm. b, Schematic of the experimental setup. c, EEG traces (n = 24) at 0.39 mW of light power, with individual (gray) and averaged (red) responses. d, Left: increased laser power leading to decreased latency time, box and whisker plots: minimum and maximum (whiskers), mean (center line) and interquartile range (box). Right: each dot represents the peak latency time averaged over n = 24 measurements (1.58 mW, 10 ms). Error bar means standard deviation (n = 4 animals). The two vertical bars were obtained by averaging the dot heights. e, Higher laser power increased LFP negative peak amplitude, whisker plots: minimum and maximum (whiskers), mean (center line) and interquartile range (box). f, EPI-fMRI and EEG recordings. n = 160 EEG traces, gray, individual; red, average. g, The S1FP region showed a modulated signal pattern highly synchronized with the stimulation and the EEG. h, The color-coded activation map overlapped on SIGMA rat brain atlas34. Single-sided GLM-based t-statistics in AFNI were used, P(FDRcorrected) = 0.0001. i, Averaged fMRI time courses from activated S1FP. Each data point in the magenta curve is represented as the mean values (center of error bar, n = 4 animals) ±s.d. (error bar). j, Averaged BOLD responses across light intensities. k, The EEG peak intensity and the maximum fMRI signal responses at 0.97 mW (blue) and 2.2 mW (magenta). l, BOLD peak responses under two laser powers (mean ± s.d., 32 data points, 8 from each animal, n = 4 animals). Opto., optogenic; stim., stimulation.

Source data

Full size image

EEG spikes were robustly recorded with laser power dependency (0.39 mW, 0.97 mW, 1.58 mW, 2.20 mW, 2.83 mW in Extended Data Fig. 2) and laser width dependency (1 ms, 5 ms, 10 ms, 15 ms, 20 ms in Extended Data Fig. 3). Increasing the laser power reduced the latency of negative peaks (Fig. 5d). When laser power was increased, the decreased latency was also observed for positive peaks that always followed the negative peaks in the EEG curve (Fig. 5d). No LFP spikes were observed in a control rat without AAV-ChR2-mCherry expression with even the highest laser power (Extended Data Fig. 4), confirming again negative peaks in Extended Data Figs. 2 and 3 to optogenetically evoked activation. Meanwhile, higher stimulation power also increased peak intensity (Fig. 5e, n = 4 animals). Notably, Rat 2 showed a larger deviation due to the electrode being placed shallower than the virus injection site, verifying the spatial specificity of optogenetic stimulation and the longer neuro-current travel time. To confirm the rat’s alertness, we also used the same method to retrieve EEG signals when the rat was stimulated by electrical forepaw stimulation (Extended Data Fig. 5).

Simultaneously, we repetitively executed the EPI sequence (in the presence of encoding gradients and RF pulses) and concurrently recorded the oscillation signal during optogenetic stimulation. Using the reconstruction algorithm (Fig. 3a and Extended Data Fig. 6), we could simultaneously retrieve EEG and fMRI signals. When the optical stimulation power was 0.97 mW, the LFP pattern (Fig. 5f) had the most peaks retaining similar amplitudes to those acquired in the absence of gradients or RF pulses (Extended Data Fig. 7a). When increasing the optical power level from 0.97 mW to 2.20 mW, the LFP peaks retrieved in the presence and in the absence of RF pulses and encoding gradients maintained with comparable intensity (Extended Data Fig. 7c,d), thus demonstrating the robustness of the wireless oscillator for LFP signal encoding during MR signal acquisition. In the concomitantly acquired EPI images, the S1FP region had an average of ~3% intensity increase during the 4-s stimulation period of 0.97 mW pulses (Fig. 5g), which was consistent with previous studies38,39,40 using conventional surface coils. This fMRI signal pattern was also synchronized with the LFP pattern (Fig. 5f).

Based on the mask defined in the color-coded activation map (Fig. 5h), we evaluated the signal intensity over multiple EPI acquisitions and consistently obtained a similar time-dependent fMRI modulation pattern (Fig. 5i,j). Plotting the maximum fMRI signal changes for each epoch against peak LFP intensity revealed an approximate linear relationship (Fig. 5k,l), demonstrating neurovascular coupling and MRI signal modulation. After imaging experiments, we removed the WISDEM and directly connected electrodes to a conventional EEG recorder. EEG signals obtained from the oscillator had a similar intensity to those obtained from a commercially available device in the absence of imaging gradients (Extended Data Fig. 8), thus once again demonstrating the reliable performance of the WISDEM. But unlike the commercially available recorder, which fails during EPI imaging due to artifact saturation, our WISDEM detector can encode and transmit EEG signals despite switching gradients. In addition, our bench tests indicate that WISDEM can reliably encode impulse frequencies up to 1 kHz (Extended Data Fig. 9) with reconstructed output amplitude comparable to input signals, sufficient for detecting multi-unit or single-unit action potentials.

Discussion

We have created a WISDEM-based platform for concurrent fMRI and EEG recording on sensory and optogenetic stimulation. This platform combines LFP and BOLD recording through the MR console system without the need for extra recording equipment or internal power source. The two-in-one WIDEM detector can wirelessly communicate with any type of signal interface that is already available on commercial MRI scanners. We have demonstrated the WISDEM detector’s ability to characterize neuronal responses during optogenetic stimulation, offering improved cellular specificity and reduced electrophysiological noises compared to electrical stimulation.

The WISDEM transducer consists of two nonlinear circuits that can be individually optimized, that is, the VSR for low-frequency signal encoding and the PR for high-frequency signal encoding and wireless carrier broadcasting. The VSR is constructed by connecting the emitter and collector terminals of a single BJT with an figure-of-eight-shaped conductor wire, unlike the previous design33 that used two varactor diodes in a head-to-head configuration. Since only two soldering junctions instead of four are required to complete the resonance circuit, the VSR has a higher quality factor (Q = 75) with smaller parasitic resistance. Another advantage of the WISDEM is that the transducer’s signal encoding capability is minimally perturbed by imaging sequences since MRI signals are received by the butterfly mode of the PR that is tuned approximately to the proton Larmor frequency (7 T). Because the PR’s butterfly mode has a magnetic field pattern perpendicular to the B1 field from the volume coil, interference is minimized, preserving signal fidelity. Unlike traditional setups requiring specialized electrodes or recording instruments to minimize noise contamination inside MRI scanners41,42, WISDEM eliminates the need for such modifications.

Previous EEG–fMRI studies relied on graphene electrodes9 to reduce electromagnetic artifacts propagating along connection cables. By eliminating connection cables, WISDEM allows greater flexibility in electrode placement, preserving space for optical fibers needed for light delivery into the brain. This also allows a more flexible choice of electrode materials, including widely used metallic electrodes with closely matched susceptibility, simplifying implementation. Furthermore, conventional MR-compatible wireless transducers25,26 required on-board gradient sensors to halt EEG recording whenever the encoding gradient switched polarity following each k-space encoding line. In contrast, WISDEM continuously encodes EEG and fMRI signals onto its oscillation carrier throughout the entire MR acquisition windows, even amid rapid gradient switching. This ensures reliable frequency-modulated-encoded signals retrieval.

The compact design of WISDEM does not incorporate additional hardware for quadrature detection, thus the MR image produced is less sensitive than a conventional cable-connected coil of the same dimension. The phase multiplication scheme is used to reconstruct a correct image in the center one-third portion of an extended FOV, which will inevitably introduce additional noise that can be considered a tradeoff for the convenience of simultaneous EEG–MRI detection. Despite its reduced SNR, the wireless multimodal WISDEM can be positioned closer to the region of interest, even within body cavities. This proximal detection scheme may still enhance focal sensitivity where a conventional surface coil is not easily accessible. Future improvements could implement quadrature frequency-modulated encoding by creating a second oscillation carrier wave with a splitter and 90° phase shifter, recovering ~29% SNR loss due to the lack of quadrature detection. Additional engineering effort is expected to minimize an extra ~10% SNR loss from fabrication imperfections.

In addition to observing focal regions near the brain surface, there are also demands to monitor neuronal activation in deep-lying regions across the entire brain. In our prototype device, the PR has a diameter of 13.5 mm, creating a butterfly mode with an effective detection depth of ~10 mm that is already comparable to the radius of a rat brain. To further enlarge the detector’s effective depth, we can potentially use the circular mode of the PR to receive MR signals from deeper regions and align the detector’s normal axis approximately perpendicular to the B1 field of a linear-mode volume coil that will be used for nuclear spin excitation. Such an arrangement can use the additional 1.4-fold SNR gain of a circular-mode detector and at the same time minimize interfering interactions from the MR excitation pulses, leaving the residual interference easily removable by a baseline correction algorithm.

For a proof-of-concept demonstration, WISDEM measured LFP signals via an 80-µm electrode. Future work will focus on integrating thinner electrodes to capture faster neuronal signals, for example, single-unit or multi-unit activities. Neuronal activities from different brain layers could be possibly detected by multiple electrodes and connecting all electrodes to the same VSR via a multiplexer43,44,45. As consequence, the same detector can consecutively encode neuronal voltages from multiple electrodes while continuously encoding MRI signals from the same FOV. To simultaneously detect multiple brain regions over an extended FOV, an array of WISDEM detectors, whose individual detector is independently manipulated by wireless activation using a unique pumping frequency46, could be constructed.

In conclusion, we have fabricated a wirelessly powered oscillator that can encode both low-frequency and high-frequency signals for simultaneous EEG and fMRI. Without the need for cable connection to a separate EEG recorder, this hybrid two-in-one transducer can be easily implemented by the neuroscience community. One preclinical application is to incorporate WISDEM into the photometry and optical-fiber-based optogenetics for simultaneous fluorescent calcium recordings and electrophysiology during optogenetic stimulation47. Clinically, this two-in-one wireless detector could benefit epileptic patients with electrodes already implanted by facilitating simultaneous imaging, thus better localizing disease foci, improving diagnostic precision and treatment strategies.

Methods

Governing equation

The PR requires an externally provided pumping signal to oscillate. Normally, the pumping frequency ωp can be experimentally adjusted by tuning the knob of an external frequency synthesizer. Once ωp is set, it will also determine the sum of butterfly and circular-mode oscillation frequencies, that is, ωp = ωb + ωc. The exact values of ωb and ωc can be derived from the following equation that describes the equal relation of reactance-to-resistance ratio in both resonance modes:

$$\begin{array}{l}\displaystyle \frac{{X}_{\mathrm{b}}}{{R}_{\mathrm{b}}}=\frac{2\left({\omega }_{\mathrm{b}}-{\omega }_{\mathrm{br}}\right){L}_{\mathrm{b}}}{{R}_{\mathrm{b}}}=2\left(\frac{{\omega }_{\mathrm{b}}}{{\omega }_{\mathrm{br}}}-1\right){Q}_{\mathrm{b}}\\ \displaystyle \quad\;\;=\frac{{X}_{\mathrm{c}}}{{R}_{\mathrm{c}}}=\frac{2({\omega }_{\mathrm{c}}-{\omega }_{\mathrm{cr}}){L}_{\mathrm{c}}}{{R}_{\mathrm{c}}}=2\left(\frac{{\omega }_{\mathrm{c}}}{{\omega }_{\mathrm{cr}}}-1\right){Q}_{\mathrm{c}}\end{array}$$
(1)

In equation (1), Lb and Lc are effective inductance of the butterfly and circular modes, while Rb and Rc are effective resistance of the butterfly and circular modes. By plugging ωb = ωp − ωc into equation (1), the butterfly-mode oscillation frequency can be calculated as:

$${\omega }_{\mathrm{b}}=\frac{{L}_{\mathrm{c}}{R}_{\mathrm{b}}{\omega }_{\mathrm{p}}+{L}_{\mathrm{b}}{R}_{\mathrm{c}}{\omega }_{\mathrm{br}}-{L}_{\mathrm{c}}{R}_{\mathrm{b}}{\omega }_{\mathrm{cr}}}{{L}_{\mathrm{c}}{R}_{\mathrm{b}}+{L}_{\mathrm{b}}{R}_{\mathrm{c}}}$$
(2)

If a VSR can somehow modulate the circular-mode resonance frequency ωcr without affecting the butterfly-mode resonance frequency ωbr, the value of ωb can also be effectively modulated.

WISDEM fabrication

To fabricate a PR, we first used a CNC milling machine to create a circuit pattern on a copper clad G10 board. This pattern consisted of a circular inductor with an inner diameter of 13.46 mm and an outer diameter of 14.46 mm, leading to an effective inductance of 29.9 nH. Within this circuit pattern, the upper and lower half circles had split gaps that were filled by varactor diodes (cat. no. BBY53, Infineon) connected in head-to-head configuration (Fig. 1a,b). The VSR was fabricated by wrapping a 32-G enameled copper wire around two 1.46-mm diameter rods that were separated by 1.8 mm (Fig. 1c). Each counterclockwise turn in the first rod was followed by a clockwise turn in the second rod. In this way, five turns with opposite orientations were wrapped around each rod before the wire’s two end terminals were connected to the emitter and collector of a BJT (cat. no. MT3S111, Toshiba), creating an effective resonance at 386 MHz (Q = 75). Meanwhile, the BJT’s base and emitter are connected via a 475-kΩ resistor (cat. no. RC3-0603-4703J, IMS) to neutralize excessive charge accumulated on the BJT’s base. The BJT’s base is connected with a sensing electrode via a 10-kΩ resistor and its emitter is connected with a grounding electrode via another 10-kΩ resistor.

The VSR was overlapping across the circular edge of the PR (Fig. 1d), with one loop sitting inside the PR and another loop sitting outside, creating an effective interaction with only the circular mode of the PR. To efficiently activate the PR at the sum frequency of its circular and butterfly modes, the pumping field was locally concentrated by an oblong shaped enhancer surrounding the PR. Fabricated out of a loop conductor with a 15.46-mm width and 20-mm length, the enhancer was empirically tuned by a trim capacitor that filled its conductor gap (Fig. 1e).

Reference coil fabrication

For sensitivity comparison, a cable-connected coil was separately built to have the same dimension as the PR described above. This cable-connected coil was made of a circular-shaped conductor with a 13.46-mm inner diameter and a 14.46-mm outer diameter (orange circle in Extended Data Fig. 1a). The circular arch was a continuous conductor, while the center conductor was split by a gap and filled with a trim capacitor (cat. no. 9410-2, Johnson) with a capacitance range from 4 pF to 18 pF. This trimmer was placed in parallel with a PIN diode (cat. no. BAR63-02V, Infineon) to create a butterfly-shaped resonance mode, enabling circuit detuning during slice excitation pulses. The right side of the gap was connected to the center pin of a coaxial cable, while the left side of the gap was connected to the outer shield of the cable via a trim capacitor (cat. no. 9702-2, Johnson) with a capacitance range from 2.5 pF to 10 pF.

Animals

All procedures in this study were approved by the Institutional Animal Care and Use Committee of Michigan State University (PROTO202200383). In total, 10 8–14-week-old Sprague Dawley rats, 5 (3 females and 2 males) for forepaw stimulation experiments, 4 (2 females and 2 males) for optogenetic-fMRI experiments and 1 female rat for control experiment without AAV-ChR2-mCherry, from Charles River were used in this study. All animals were three-in-one-housed in 12/12 h on/off light/dark cycle conditions to assure undisturbed circadian rhythm and ad libitum access to chow and water.

Optogenetic virus injection

AAV5.CaMKII.hChR2 (H134R)-mCherry was purchased from Addgene (no. 26975-AAV5) and packaged into frozen-stored vials, each of which contained 100 µl of sample at a concentration higher than 1 × 1013 vg ml−1. To inject AAV5 into the right somatosensory forepaw region of brain in a 4-week-old rat, the rat was first anesthetized with 5% isoflurane in an induction box, then secured on a stereotaxic frame and maintained at 1.5–2% isoflurane via a nose cone. An incision was made on the scalp to expose the skull before craniotomies were made with a pneumatic drill to introduce minimal damage to cortical tissue. Afterward, 0.4–0.6 µl of viral droplet was injected from a 10-µl syringe via a 35-gauge needle to the following coordinates: 0 mm posterior to the Bregma, 3.2–3.5 mm lateral to the midline and 0.5–1.2 mm below the cortical surface using an infusion pump (Pump 11 Elite, Harvard Apparatus). Once the AAV5 injection was finished, the needle was left in place for approximately 5 min before being slowly pulled out. The craniotomies were sealed with the bone wax, and the skin around the wound was sutured. After the surgery, rats were subcutaneously injected with antibiotics and painkillers (Ketoprofen fluids) for three consecutive days to prevent infection and to relieve pain. Imaging experiments were performed 4 weeks after virus injection to allow enough ChR2 protein expression in the S1FP region.

Animal preparation, electrode and fiber implantation

To prepare for electrode implantation, an enameled copper wire of 80-µm diameter was glued with an optical fiber of 200-µm core diameter (cat. no. FT200EMT, Thorlabs). The front edge of the copper wire was trimmed by a scissor to expose the conductor tip for direct contact with the brain tissue. Once the animal was anesthetized by 2–2.5% isoflurane with its head fixed on a stereotaxic frame, a burr hole of 1.5-mm diameter was drilled on the rat’s skull so that the dura could be carefully removed. Afterward, the optical fiber along with the attached electrode was inserted into the S1FP region, at coordinates of 0 mm posterior to Bregma, 3.2–3.5 mm lateral to the midline and 1.2 mm below the cortical surface. Subsequently, an adhesive gel (Loctite 454, Henkel) was applied on the insertion hole to secure the fiber–electrode assembly against the skull. A grounding electrode was separately screwed against the skull above the neck. After the scalp was closed by glue, the rat was injected by a bolus of dexmedetomidine (0.05 mg kg−1, subcutaneously; Dexdomitor, Orion Pharma) before isoflurane was discontinued. The rat was then transferred into the MR scanner with its head secured by a bite bar and two ear bars, so that the WISDEM detector could be mounted above the rat’s head to cover the right S1FP region. Finally, both the sensing electrode and the grounding electrode were connected to the plug-in pins that were previously soldered on the VSR.

During in vivo imaging, the rat was subcutaneously administered with a constant infusion of dexmedetomidine at 0.1 mg kg−1 h−1. Its breathing and heart rates were monitored by an air pillow placed beneath its chest that was interfaced with an MR-compatible monitoring system (Model 1025, SA Instruments, Inc.). The rat’s body temperature was continuously monitored by a rectal probe (SA Instruments) and maintained at 37 °C by a water jacket.

fMRI acquisition

All MRI data were acquired inside a 7-T small-animal scanner (Bruker BioSpin) with a 16-cm horizontal bore. The WISDEM detector was placed above the animal’s skull and activated by a loop antenna to produce a sustained oscillation signal. Functional MR images were acquired with a multi-slice gradient-echo EPI sequence with the following parameters: echo time (TE) = 20.381 ms, repetition time (TR) = 0.9976 s, FOV = 78 × 26 × 14.4 mm3, matrix size 129 × 43, voxel size 0.6 × 0.6 mm2, slice number 24, slice thickness 0.6 mm, flip angle 90° and bandwidth 326.087 Hz. Concurrently during image acquisition, we also performed electrical stimulation on the rat’s left forepaw or optogenetic stimulation in the S1FP region. For both the forepaw and optogenetic stimulation, the stimulation paradigm started from a 12-s prestimulation delay, followed by eight epochs of stimulation cycles, each of which started from a 2-s resting period followed by 4-s stimulation period and concluded with a 9-s interval. The total repetition number for the entire EPI experiment was 135. The 4-s period for electrical forepaw stimulation contained 20 biphasic pulses, each of which had a 333-μs duration and 5-Hz repetition rate. The 4-s period for optogenetic stimulation contained 20 square pulses, each of which had a 10-ms duration and 5-Hz repetition rate.

MR image reconstruction

Once the MR scanner recorded the complex oscillation signal from the WISDEM, its instantaneous phase angle was obtained by the angle and unwrap functions in MATLAB. The phase derivative \({{\mathrm{d}}\varnothing }_{t}/{\mathrm{d}t}\) at each time point was obtained by multiplying the scanner’s sampling speed (326,087 Hz) with the phase difference between adjacent time points. This phase derivative \({{\mathrm{d}}\varnothing }_{t}/{{\mathrm{d}}t}\) was high-pass filtered by subtracting \({{\mathrm{d}}\varnothing }_{t}/{{\mathrm{d}}t}\) with its smoothed value using the formular \({\mathrm{HPF}}\left({{\mathrm{d}}\varnothing }_{t}/{{\mathrm{d}}t}\right)={{\mathrm{d}}\varnothing }_{t}/{{\mathrm{d}}t}-{\mathrm{smooth}}({{\mathrm{d}}\varnothing }_{t}/{{\mathrm{d}}t},8)\) so that low-frequency baseline variation due to EEG signals were removed (Extended Data Fig. 6a). The intense spikes that appeared between adjacent slices were due to discontinuous phase values before and after the blanking periods between adjacent acquisition slices. These blanking periods were reserved for data transfer and spectrometer reset. Because the intense spikes would occur only once at the beginning of each acquisition slice, they occupied a small portion of time and could simply be discarded (Extended Data Fig. 6b). To maximize the sampling time of EEG signals during image slice acquisition, each k-space line was sampled during both positive and negative gradients. The retrieved k-space signals for the positive (Extended Data Fig. 6c) and negative (Extended Data Fig. 6d) gradients were separated and multiplied by the phase term \(\exp ({-j\varnothing }_{t})\) of the oscillation signal, before 2D Fourier transformation was performed on both datasets to obtain two images that were combined into a single frame (Extended Data Fig. 6e).

EEG signal reconstruction

The phase derivative \({{\mathrm{d}}\varnothing }_{t}/{{\mathrm{d}}t}\) was low-pass filtered below a 1-kHz passband, leading to a raw EEG signal (Extended Data Fig. 6f) containing a repetitive baseline variation that was synchronized with slice acquisition. This average baseline pattern had 43 pairs of positive and negative peaks, corresponding to 43 k-space lines, each of which contained one positive gradient slope and one negative gradient slope. To flatten the baseline, the reconstructed signal corresponding to each slice was subtracted by the average baseline pattern (Extended Data Fig. 6g), before the median filter and linear interpolation were applied to remove intensity spikes due to blanking periods between adjacent slices (Extended Data Fig. 6h–k). For comparison purposes, EEG signals were also recorded by a commercial recorder (Biopac MP160) with a direct wire connection to the sensing electrodes.

fMRI data analysis

After functional images were retrieved using the algorithm described in Fig. 3a and Extended Data Fig. 6, they were imported into the AFNI software (analysis of functional neuroimages, National Institutes of Health) for subsequent analysis. First, functional images were aligned to anatomical images that were separately acquired with a rapid-acquisition-with-relaxation-enhancement sequence. These anatomical images had the same orientation and FOV as functional images but with a higher spatial resolution at 100 µm. The anatomical images were then registered to SIGMA rat brain template34 to derive the transformation relation, based on which the functional EPI images were registered to the same standard template. The baseline signal of EPI images was normalized to 100 for statistical analysis of multiple trials over the time course. The time courses of the BOLD signal were extracted from the S1FP region because this region had notable activation values in the brain atlas. The hemodynamic response function used the BLOCK function in the linear program 3dDeconvolve, where the statement BLOCK (L, 1) was a convolution of a square wave of duration L and peak amplitude of 1. To compute the evoked BOLD changes in Figs. 4d and 5h, 3dmaskave was evaluated in the region of interest defined as the primary somatosensory area in the SIGMA atlas34.

Immunohistochemistry

To verify the phenotype of the transfected cells, opsin localization and optical-fiber placement, the rat was euthanized and perfused in its left ventricle. The rat brain was extracted, fixed overnight in 4% paraformaldehyde and then equilibrated overnight in 15% sucrose dissolved in 0.1 M phosphate buffer at 4 °C, before being soaked inside 30% sucrose dissolved in 0.1 M phosphate buffer. Subsequently, the brain was sectioned to 30-µm slices on a sliding microtome (Leica CM 1850). Free-floating brain slices were washed in PBS, mounted on microscope slides and incubated with 4,6-diamidino-2-phenylindole (Sigma Aldrich) at room temperature, before being imaged by a fluorescence microscope (Nikon A1 Laser Scanning Confocal Microscope, NIS-Elements AR v.4.50) for assessment of ChR2 expression in the S1FP region (Fig. 5a). To enhance brightness and contrast for visualization purposes, digital images were minimally processed using ImageJ for better visualization.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.