PAM system
The system configuration is shown in Fig. 1a. The excitation source is a nanosecond-pulsed laser (PAM laser; VGEN-G-20, Spectra-Physics). A half-wave plate (HWP; WPH10M-532, Thorlabs), an electro-optic modulator (EOM; 350-50-01-RP, Conoptics), and a polarizing beam splitter (PBS; PBS121, Thorlabs) work together as a high-speed switch to distribute individual laser pulses between two optical paths. When a low voltage is applied to the EOM, the laser pulse passes through the PBS. Then, the energy of the pulse is adjusted by a second HWP (WPH10M-532, Thorlabs) and a second PBS (PBS121, Thorlabs) before being coupled into a 10-meter-long polarization-maintaining single-mode fiber (PM-SMF; HB450-SC, Fibercore) for wavelength conversion based on the SRS effect. Finally, a band-pass filter (BPF; ZET561/10X, Chroma) is applied to the output of the PM-SMF to isolate the 558-nm component. When a high voltage is applied to the EOM, the polarization of the laser pulse is rotated by 90°. As a result, it is reflected by the PBS to a different path, where no wavelength conversion occurs. The laser pulses coming out of the two paths have different wavelengths (i.e., 558 nm and 532 nm), which are combined using a dichroic mirror (FF552-Di02-25 × 36, Semrock) and coupled into a single-mode fiber (SMF; P1-460B-FC-1, Thorlabs). Before the SMF, a beam sampler (BS; BSF-05-A, Thorlabs) and a photodetector (PD; PDA10A2, Thorlabs) are used to record the temporal jitter and energy fluctuation of the laser pulses for offline compensation. The dual-wavelength output of the SMF is scanned by a 2D galvanometer (GM; GVS202, Thorlabs) and relayed by a lens pair (L1; AC508-80; L2; AC508-400, Edmund optics). A second nanosecond-pulsed laser performs the single-vessel occlusion (stroke laser; GLPM-10, IPG Photonics). Two HWPs (WPH10M-532, Thorlabs) separately adjust the polarization states of the laser pulses used for PAM imaging and vessel occlusion, and a PBS (PBS121, Thorlabs) combines the two types of pulses with orthogonal polarization states. The combined beam is then focused by an objective (MY-5X-802, Thorlabs) through the transparent MRR to the target to be imaged. A tunable CW laser (MRR laser; TLB-6712, Newport) is used for MRR interrogation, and a manual polarization controller (PC; FPC022, Thorlabs) adjusts the polarization state of the interrogation light. The output of the MRR is detected by a balanced photodetector (BPD; PDB465A-AC, Thorlabs) and acquired by a field-programmable gate array (FPGA; PCIe-7842, National Instruments). The DC output of the MRR is input into a closed-loop control in LabVIEW (see the “Proportional-integral-derivative (PID) control” section below for details), while the radio-frequency output is digitized by a high-speed waveform digitizer board (DAQ; ATS9350, AlazarTech) for PAM image formation.
Micro-ring resonator
An 80-µm-diameter polymer MRR and a matching waveguide are fabricated on a 250-µm-thick quartz plate using the soft nanoimprinting lithography method, similar to that previously reported26. A 6-µm-thick polymer layer (MY-131-MC, MY Polymer) is spin-coated to fully encapsulate the MRR on the quartz substrate. Then, the MRR is coupled to an SMF (S630-HP, Thorlabs) as the input port and a multi-mode fiber (MMF, GIF625, Thorlabs) as the output port, on an 8-mm-diameter glass coverslip. In comparison to the previous MRR26, the MRR used in this study features two major design innovations to improve its sensitivity for ultrasonic detection. First, we have changed the cladding materials from polydimethylsiloxane (PDMS) to MY-131-MC, which has a lower refractive index of 1.31. It results in an improved Q-factor of the MRR (from 4.2 × 104 to 1.3 × 105) and thus higher ultrasonic detection sensitivity. Second, we have improved the coupling efficiency from the SMF to the MRR by integrating an adiabatically tapered waveguide. The wider opening of the tapered waveguide reduces its mode mismatch with the SMF and improves the coupling efficiency from 5 to 19%, which results in an increased signal-to-noise ratio of recorded photoacoustic signals. More detailed information about the design and fabrication of MRR can be found in the Supplementary Information.
Proportional-integral-derivative control
Laser-induced thermal accumulation during PAM imaging can shift the resonant frequency of the MRR, thereby impairing its sensitivity. To address this issue, a closed-loop PID controller is used to dynamically lock the interrogation light at the MRR’s resonant wavelength. First, the wavelength of the MRR laser is set to the MRR’s resonant wavelength. Then, the PID control module in the customized LabVIEW program is activated. The fast-monitoring output of the BPD (i.e., the DC output of the PDB465A-AC) is captured by the FPGA and fed into the PID control module, based on which a feedback analog signal is generated by the DAQ to fine-tune the MRR laser wavelength. By optimizing the PID parameters (typically, proportional gain: 0.0005; integral gain: 0.005; derivation gain: 0.05; filter coefficient: 0.2), the DC output from the MRR can be locked at a point where the sensitivity is maximized by tuning the interrogation wavelength of the MRR laser. Time-lapse photoacoustic signals of a black tape with and without PID control are shown in Fig. 1d, showing that the feedback control effectively maintains the MRR’s sensitivity at the optimal level.
System characterization
The original lateral and axial resolution of the PAM system is characterized using a resolution target (R1DS1P, Thorlabs). Specifically, the edge spread function was experimentally measured by scanning the light focus across the sharp edge of a bar pattern on the target (the red curve in Fig. S3a), based on which the line spread function was derived (the pink curve in Fig. S3a). Then, the lateral resolution was estimated to be 3.2 µm, based on the full-width-at-half-maximum (FWHM) value of the line spread function. The axial resolution was estimated to be 12 µm, based on the FWHM value of an A-line photoacoustic signal of the resolution target after Hilbert transform (Fig. S3b). Note that the axial resolution significantly degraded in vivo due to the strong attenuation of high-frequency ultrasound in tissues, as evidenced in Fig. 4. The cross-sectional profile of L2 shows that conventional PAM completely fails to resolve the two microvessels along the axial direction.
Scanning schemes
The high frame rate provided by the PAM system relies on the high-speed laser scanning within the acoustic FOV of the MRR. Our PAM system has two different scanning schemes: 2D real-time imaging (Fig. 2 and Movie S1) and 3D super-resolution imaging (Figs. 3–6 and Movies S2–S4). Specifically, to achieve 2D real-time imaging (i.e., RT-fPAM), the PAM laser is operated at 150 kHz for 2D galvo scanning over a 150 × 150 µm2 region with a frame rate of 15 Hz (each frame contains 100 × 100 pixels). To achieve 3D super-resolution imaging (i.e., SR-fPAM), the pulse repetition rate of the PAM laser is set to 200 kHz, and the galvanometer’s line rate is set to 1 kHz. Before advancing to the next B-scan position, 800 repeated B-scans are acquired for RBC tracking. In addition, to achieve blood flow quantification in 3D, two series of B-scans along both lateral directions (i.e., X and Y) are collected, and RBC tracking is performed on each of them and finally combined.
Functional imaging
We have previously reported simultaneous quantification of microvascular blood oxygenation (sO2) and flow using spectroscopic and correlation analyses in conventional PAM27,28. In this study, sO2 is measured as before, which employs dual-wavelength (i.e., 532 nm and 558 nm) laser pulses for excitation to distinguish oxy- and deoxy-hemoglobin based on differences in their optical absorption spectra. It is worth noting that the SRS output at 558 nm exhibits high stability (fluctuation <2%), and the power fluctuations are compensated by normalizing the A-line signal to the PD-recorded light intensity. As for blood flow speed quantification, 2D or 3D tracking of RBCs is performed. Detailed methods for the flow quantification are described below.
Flow quantification by Hough transform in RT-fPAM
In RT-fPAM, the high volumetric imaging rate (i.e., 15 Hz) enables RBC tracking in sequentially acquired en-face images. As shown in Fig. S6a, a kymograph can be generated for a particular vessel of interest. To quantify the blood flow speed, we first normalize the kymograph and extract RBC trajectories via edge detection to form a binary image (Fig. S6b). Then, the Hough transform54 is applied to individual sliding windows (3 s × 100 µm) in the binary image to quantify the flow speed at a desired time and location. The horizontal axis (\(\theta\)) of the resulting Hough transform matrix represents the slope of the RBC trajectory (Fig. S6c). Next, the ten matrix elements with the highest intensity values are identified (highlighted by the green box in Fig. S6c), and the mean value of their corresponding angles is calculated to estimate the mean slope of RBC trajectories. Finally, the blood flow speed in the vessel of interest is quantified based on its inverse relationship with the slope of the RBC trajectory.
Red blood cell tracking for 3D functional imaging in SR-fPAM
To achieve SR-fPAM, we are inspired by the ultrasound localization microscopy20. Hundreds of B-scans are acquired with an extremely high frame rate (i.e., 1 kHz), and RBC movements in them are tracked to realize super-resolution vascular imaging. Fig. S7 shows the step-by-step flow diagram of the RBC tracking algorithm that enables 3D label-free super-resolution imaging of microvasculature and vascular functions.
Firstly, each B-scan is Hilbert-transformed after being compensated for jitter using the PD data. Then, motion and reverberation artifacts are removed from individual B-scans. Next, the pre-processed B-scans are rescaled using bicubic interpolation, and the correlation between the interpolated B-scans and the 2D point spread function (PSF) of the PAM system is performed, with the assumption that the RBC centroids are highly correlated with the PSF. The PSF is generated based on experimentally measured spatial resolutions of the MRR-based PAM system (3.2 µm laterally and 12 µm axially, respectively, as shown in Fig. S3). To suppress noise, in the 2D correlation coefficient map of each frame, coefficients below 0.35 or falling outside the vascular regions are excluded. The vascular regions are identified based on the signal intensity of the raw PAM images averaged over 800 frames, where pixels with normalized intensities greater than 0.02 are classified as vessel areas. After the correlation coefficient map is thresholded, the RBC centroids are identified by detecting local maxima. To pair centroids corresponding to the same RBC or RBC cluster in adjacent frames, a nearest-neighbor matching strategy is applied. Specifically, for each centroid, all candidate centroids in the next frame within a predefined search range are evaluated, and the one with the minimum spatial distance is selected and linked. Given the dense distribution and average size of RBCs, the maximum search range for centroid linking is set to 6 µm in both the lateral and axial directions. To further reduce misconnections, only RBC centroids that can be continuously tracked and move in a consistent direction for at least 3 ms (i.e., three successive frames) are retained. Trajectories are then formed by linking the positions of RBC centroids across adjacent frames, enabling the calculation of flow velocity from the centroid displacement and the frame rate. Finally, the structural image of the microvasculature and the blood flow map can be generated based on the number of trajectories and the average flow speed measured at each pixel, respectively.
Recognizing that RBC tracking in repeated B-scans only allows the quantification of flow velocity within the 2D imaging plane, bi-directional scanning along both lateral directions (i.e., X–Z and Y–Z) is necessary to collect orthogonal B-scan series along both lateral directions. From the X–Z B-scans, we obtain the velocity components along the X and Z axes (i.e., \({v}_{X}(x,y,z)\) and \({v}_{Z}(x,y,z)\)). For vessels that are not fully parallel to the X–Z plane, the lateral flow velocity contains a Y-direction component (i.e., \({v}_{Y}(x,y,z)\)), which cannot be recovered from X–Z scans alone and is instead obtained from RBC tracking in the Y–Z B-scans. After obtaining the three orthogonal velocity components, they are combined by vector addition in 3D (i.e., \({v}_{3D}\left(x,y,z\right)=\,\sqrt{{{v}_{X}(x,y,z)}^{2}+{{v}_{Y}(x,y,z)}^{2}+{{v}_{Z}(x,y,z)}^{2}}\)). Owing to our unique system design, changing the scanning direction is simple and straightforward: the fast-scanning axis of the galvanometer is switched by adjusting the driving voltage, and no modification of the acoustic sensor is required. Figure 5b shows the flow velocity maps along all X, Y, and Z directions, which are combined into the 3D flow speed distribution shown in Fig. 5c and Movie S4.
As for the blood oxygenation measurement, information obtained through the dual-wavelength excitation is incorporated to produce a 3D visualization of the vascular sO2, since we can quantify the sO2 inside the super-resolved 3D microvasculature. 2D projections and 3D visualization of the microvascular sO2 are shown in Fig. 5a, c and Movie S4.
Dwell time optimization for RBC tracking
Increasing the dwell time for repeated B-scan acquisitions allows more RBC trajectories to be recorded, but the associated time cost also increases. Thus, the dwell time needs to be optimized for the best tradeoff between image reconstruction fidelity and time expense. To this end, a small region in the mouse cortex was imaged with different dwell times, each corresponding to a different number of repeated B-scans for tracking (Fig. 3e, f and Movie S2). Microvascular diameters of four representative vessels (L1–L4; indicated by the white arrows in Fig. 3e) generated using different dwell times were calculated and compared with the diameter generated using the longest dwell time (i.e., 3000 frames). As shown in Fig. 3f, once the number of repeated B-scans reaches 800 (i.e., 0.8 s dwell time), further increasing the frame number for tracking does not yield a significant change (<5%) in the measured diameter. This optimization study suggests that a 0.8 s B-scan dwell time is sufficient for RBC tracking-based super-resolution image reconstruction and thus is used in SR-fPAM.
Validation of the localization-based blood flow quantification
A phantom experiment was performed to assess the accuracy of the tracking-based flow speed measurement. In this experiment, a syringe pump (NE-300, Pump System Inc.) was used to pump 3-µm-diameter black polystyrene microspheres (24292-15, Polysciences) through a transparent plastic tube at flow speeds ranging from 0.05 to 6 mm/s. As shown in Fig. S8, the flow speeds measured using the spatiotemporal RBC tracking show a good linear relationship (R2 = 0.999) with the preset values across the entire flow range.
Two-photon microscopy
To benchmark SR-fPAM, we have performed a side-by-side comparison of SR-fPAM and TPM since TPM can serve as a gold standard for high-resolution vascular imaging. A TPM image of the same mouse brain region is acquired using a homemade TPM system with a 1.05 NA objective lens (XLPLN25XWMP2, Olympus). The TPM can be easily integrated into the PAM system and utilized to collect image data of the same FOV on the mouse brain due to the optical transparency of MRR. During TPM imaging, fluorescein isothiocyanate-dextran (FITC; FD-2000S, Sigma) was intravitreally injected to label the vasculature. The brain region was scanned with a lateral step size of 0.5 µm and an axial step size of 0.625 µm for volumetric two-photon imaging. Side-by-side comparisons of the 2D projection images and 3D renderings acquired using conventional PAM, SR-fPAM, and TPM are shown in Fig. 4a and Movie S3, respectively. Quantitative assessments were performed on the cross-sectional profiles of selected microvessels (Fig. 4b) and each plane of the 3D microvasculature (Movie S3), both illustrating high similarity between SR-fPAM and TPM.
Depth compensation
In PAM, depth information is obtained based on time-resolved ultrasonic detection. Due to the small size, MRR can be regarded as a point detector, so both the depth and lateral distances between the photoacoustic source and the MRR contribute to the time delay. Leveraging the geometric relationship shown in Fig. S9a, we developed a delay compensation algorithm to accurately retrieve the true depth information. To evaluate its performance, we acquired the B-scan images of a piece of flat black tape at three different depths. Before the compensation, all three B-scans showed arc-shaped delays, with the curvature diminishing as the black tape was positioned further away from the MRR (top row of Fig. S9b). After compensation, the additional delay arising from the lateral displacement was successfully eliminated (bottom row of Fig. S9b).
Single-microvessel occlusion
We used short laser pulses to directly generate blood clots in targeted microvessels without the aid of photothrombotic dye29. Specifically, a 5-s pulse train (150 kHz repetition rate, 300 nJ energy, 1.5 ns width) was launched onto a microvessel near the center of the FOV. The location of the stroke is precisely controlled by the laser induction. Disappeared photoacoustic signals at and around the occlusion location indicate a successful occlusion (Figs. 2c, d and S10, and Movie S1; successful occlusion was repeated in five different animals).
Crosstalk elimination
The nanosecond laser pulses used to induce micro-strokes have relatively high energy, which can excite considerable photoacoustic signals. By synchronizing the two lasers used for imaging and occlusion (i.e., PAM laser and stroke laser in Fig. 1a), we can precisely control the timing of the two types of photoacoustic signals and the delay between them. With proper time-gating, the crosstalk induced by the stroke laser can be completely removed from each B-scan, as shown in Fig. S10.
Animal preparation
C57BL/6J mice (male, 9–10 weeks, The Jackson Laboratory) were used for all in vivo experiments herein. Following craniotomy, a 100-µm-thick acrylic film (Emco Industrial Plastics) was used to seal the open-skull window. After the procedure was completed, the mice were returned to their home cages and administered buprenorphine at a dosage of 0.1 mg/kg for pain management. Throughout the post-operative period, the mice were monitored for signs of distress or pain. Two weeks later, imaging experiments were performed under general anesthesia with 1.5% isoflurane, and the mice were kept at 37 °C using a heating pad. All experimental procedures were carried out in conformity with the animal protocol approved by the Institutional Animal Care and Use Committee at Washington University in St. Louis.