Three-dimensional aperture-driven photoacoustic tomography (3D-ADPAT)

Xuanhao Wang; Yang Xiao; Yuqi Wang; Xiali Gao; Yongdu Ruan; Xiaohui Yang; Yiyin Su; Dikui Zhou; Xiangdi Li; Yuqian Meng; Zhibo Xiao; Fan Meng; Ruofan Wang; Junhui Shi

doi:10.1126/sciadv.aeb7362

. 2026 Jun 26;12(26):eaeb7362. doi: 10.1126/sciadv.aeb7362

Three-dimensional aperture-driven photoacoustic tomography (3D-ADPAT)

Xuanhao Wang ^1,^2,^†, Yang Xiao ^1,^2,^†, Yuqi Wang ^1,², Xiali Gao ³, Yongdu Ruan ^1,², Xiaohui Yang ^1,², Yiyin Su ^1,², Dikui Zhou ^1,², Xiangdi Li ^1,², Yuqian Meng ^1,², Zhibo Xiao ³, Fan Meng ³, Ruofan Wang ^1,², Junhui Shi ^1,^3,^2,^*

PMCID: PMC13308596 PMID: 42361176

Abstract

As a powerful modality for biomedical research, photoacoustic computed tomography (PACT) offers the combination of optical specificity and acoustic penetration for in vivo visualization. However, its performance is limited by a trade-off between ultrasound detection aperture and sensitivity, compromising image contrast, resolution, and penetration depth. Here, we introduce three-dimensional aperture-driven photoacoustic tomography (3D-ADPAT) to overcome these limitations, synergistically combining a high–numerical-aperture focused ultrasonic transducer (HNA-FUT) with a phase-inverted focusing-equivalent reconstruction (PIFER) algorithm. This approach yields ~2.21-fold enhancement in spatial resolution and ~10-fold increase in contrast-to-noise ratio. 3D-ADPAT’s performance was validated through comprehensive simulations and phantom studies and further confirmed by in vivo whole-body imaging that resolved deep-seated anatomies and tracked small-molecule metabolic pathways. Furthermore, a barrel-shaped array–based 3D-ADPAT implementation was designed, supporting dynamic imaging within a large field of view or across variable spatiotemporal scales. In summary, 3D-ADPAT provides a powerful tool for advanced photoacoustic systems, promising to accelerate frontier biomedical research.

A 3D photoacoustic tomography system breaks a fundamental bioimaging trade-off, revealing deep tissues with high clarity.

INTRODUCTION

Photoacoustic (PA) imaging uses short-pulsed laser excitation to convert optical absorption in biological tissues into ultrasonic signals, thereby achieving imaging with both optical contrast and acoustic resolution (1). As the only noninvasive modality capable of mapping optical absorption at centimeter-scale depths in living organisms, three-dimensional (3D) photoacoustic computed tomography (PACT) synthesizes acoustic apertures through transducer arrays or element scanning to reconstruct the distribution of diffused light energy in deep tissues (2, 3). Previous theoretical studies on 3D PACT sampling have analyzed optimal transducer array configurations and element densities required for resolvable fields of view (FOVs), all based on the assumption of ideal point detectors (4). Such detectors are conceptually defined as entities with dimensions smaller than half the central wavelength, capable of receiving signals over a full 4π steradian solid angle without noise interference (5). However, practical implementations in 3D PACT arrays deviate from this ideal model. For example, in spherical arrays, transducer elements degenerate into planar transducers (PTs) where finite-aperture effects cause a radial deterioration of resolution from the center, effectively integrating the ideal reconstruction over a degraded aperture surface (6, 7). In ring arrays, anisotropic element apertures lead to substantial degradation of transverse resolution along the scanning direction (8, 9). Practically, finite-aperture transducers exhibit limited reception solid angles, necessitating high detection density (10). While larger element sizes ensure high signal-to-noise ratio (SNR), the detected amplitude scales proportionally with detector area, whereas thermal noise scales inversely with it, resulting in critically low SNR for quasi-point transducers (11).

Several studies have attempted to address finite-aperture effects through reconstruction algorithms. Modified back-projection (12) or model-based (13) methods incorporate detector surface modeling to improve upon point-detector assumptions, enabling more accurate reconstructions (14, 15). However, these approaches often demand substantial computational resources and have primarily been validated in simulations or simple phantoms. Alternative algorithms mitigate aperture-induced blurring by deconvolving calibrated spatial impulse responses, yet these methods are highly noise-sensitive and introduce calibration errors, failing to deliver satisfactory results in in vivo applications (16, 17). Crucially, algorithmic modifications only optimize the use of finite-aperture data without recovering information from acoustically unsampled regions. Although deep learning has advanced PA imaging, no dedicated solution addressing this fundamental limitation has been reported to date (18, 19).

Parallel hardware-focused research aims to develop ultrasound probes with high SNR and wide acceptance angles. All-optical ultrasound transducers can approach ideal point detection but face challenges in fabrication complexity, stability, and scalable array integration (20–22). Other researchers have fabricated piezoelectric ring detectors that wrap around specimens, using the ring center as a virtual reception point for scanning (23). While partially achieving large apertures and high SNR, this approach lacks practicality due to mechanical constraints (24). Commercial concave/convex probes or acoustic lens transducers combined with virtual-point reconstruction have expanded the FOV of 2D PACT, but unoptimized reconstruction methods result in poor axial focusing, low contrast-to-noise ratio (CNR), and unverified feasibility for 3D PACT (25–28). Mechanical secondary scanning (element rotation around its axis) enlarges detection apertures but is incompatible with rapid volumetric acquisition and impractical for dense 3D PACT arrays (29). Traditional line-focused arrays with virtual-point reconstruction offer limited improvement for 3D PACT due to small numerical apertures (NAs) (30). Slit-enhanced linear arrays (with slit widths near the central wavelength) exploit acoustic diffraction to improve 3D resolution isotropy, but at the cost of severe CNR degradation that compromises PACT’s depth advantage (31–34).

Here, we develop, characterize, and validate 3D aperture-driven photoacoustic tomography (3D-ADPAT), an optical imaging technology based on high–numerical-aperture acoustic detection, using specialized ultrasonic transducers with large detection solid angles and high sensitivity. Through rigorous simulations, we establish a phase-inverted focusing-equivalent reconstruction (PIFER) specifically tailored for the designed transducers to achieve high-quality 3D imaging. Comprehensive numerical simulations, phantom experiments, and in vivo anatomical and functional imaging demonstrate that, compared with conventional PACT systems, 3D-ADPAT provides 2.21-fold higher spatial resolution and 10-fold greater CNR, with substantial improvements in imaging fidelity and penetration depth. Leveraging the hardware simplicity and integrability of 3D-ADPAT, we further explore its potential for high-density 3D array designs to achieve large-FOV, high–spatiotemporal-resolution imaging. Detailed technical specifications, extensive comparative experiments, and in-depth optimization explorations are all comprehensively presented. These results collectively validate the high-fidelity imaging capabilities and broad applicability of 3D-ADPAT, establishing a pathway for PA system design and providing a powerful, accessible imaging platform for life science research.

RESULTS

Concept and system implementation of 3D-ADPAT

The design of the ultrasonic transducer is central to the imaging performance of 3D-ADPAT. For a disk-shaped transducer of diameter D, its angular sensitivity distribution in the far field is given by (35)

S_{λ} (θ) = 2 |\frac{J_{1} [π D sin (θ) / λ]}{π D sin (θ) / λ}|

(1)

where J₁ is the first-order Bessel function and λ is the central wavelength of the transducer. As governed by Bessel function properties, a large D relative to the wavelength (e.g., PTs) yields a small acceptance angle θ, corresponding to a limited reception aperture that causes information loss at fixed detection positions (Fig. 1A). While transducers with dimensions approaching the central wavelength approach an ideal aperture, their small size proportionally reduces the amplitude of received acoustic signals (Fig. 1B). Conversely, the equivalent root mean square (RMS) noise voltage v_RMS arising from thermal noise is expressed as

v_{RMS} = \sqrt{\frac{k T_{0}}{C_{T}}}

(2)

where k is Boltzmann’s constant and T₀ is temperature. Because the element area is proportional to C_T, the equivalent noise voltage scales inversely with element size. Consequently, the SNR of the probe diminishes rapidly with reduced detector size. The use of large-scale detectors with diffraction slits reconstructs the wavefront via Huygens’ principle, still resulting in critically low SNR despite achieving a high NA (Fig. 1C).

In 3D-ADPAT, we developed a high–numerical-aperture focused ultrasonic transducer (HNA-FUT) based on conventional piezoelectric ceramic (PZT) materials, designed for array integration. The HNA-FUT features a strongly focused geometry, with a focal length of 15 mm, a lateral diameter of 14 mm, indicating the NA value ~0.423 (fig. S1 and Materials and Methods for details). This configuration achieves substantial detection aperture at its focus while maintaining high sensitivity due to its large surface area (Fig. 1D). Balancing imaging depth and resolution, the central frequency was selected at 3.5 MHz with a unidirectional bandwidth exceeding 80%. To visualize the detector’s sensitivity profile, we characterized the emitted acoustic field of the HNA-FUT using an ultrasonic beam analyzer. Under sinusoidal voltage excitation at the center frequency, the amplitude distribution along the transducer’s central axis was captured (Fig. 1E). By acoustic reciprocity, the emission field qualitatively represents the reception sensitivity distribution. The far-field amplitude distribution converges toward that of an ideal point detector located at the focus (conditions for phase equivalence will be detailed later). To quantify the acceptance solid angle, conjugate spherical wavefronts centered on the focus were analyzed at multiple distances. Gaussian fitting of the angular amplitude (sensitivity) distribution yielded the full width at half maximum (FWHM). For instance, at 21 mm from the focus, the calculated half-angle was 26.2° (Fig. 1F), whereas, at 37 mm, it was 23.7° (Fig. 1G). Averaging measurements across 10 positions yielded an acceptance angle of ±25.6° for the HNA-FUT. Solving Eq. 1 inversely indicates that this acceptance angle corresponds to an equivalent point detector diameter of ~1.16 mm, a size that is challenging to implement in practical imaging systems while maintaining adequate SNR. In 3D-ADPAT, the large acceptance aperture of the HNA-FUT enhances the spatial resolution of the synthetic-aperture reconstruction, while its high SNR ensures deep penetration, which, when combined with a finely designed reconstruction algorithm, yields outstanding 3D imaging performance.

3D-ADPAT is implemented via matrix scanning of HNA-FUT elements (Fig. 1H). For operational convenience, the probe is mounted on a 2D translation stage beneath the sample stage via a linkage arm. The sample stage, fabricated from 2.5% agar, provides structural support while maintaining optimal acoustic and optical transparency. To leverage the large detection aperture, the HNA-FUT focus is positioned ~62 mm below the upper surface of the sample stage (equivalent to the probe surface being positioned at ~77 mm from the stage surface). This ensures coverage of a 60-mm-diameter sample area (e.g., small animal torso) within the ±25.6° acceptance angle while minimizing high-frequency loss and energy dissipation. A fundamental Neodymium-doped Yttrium Aluminum Garnet (Nd:YAG) laser and a solid-state laser with a frequency-doubling module serve as excitation sources, both operating at 10-Hz pulse repetition rate to enable independent 1064- and 532-nm output. The two laser beams are coaligned by a dichroic mirror and delivered via a bifurcated quartz fiber bundle. The output beams from the line-shaped fiber terminals are symmetrically directed onto the sample stage by adjustable silver mirrors, with the agar providing further light homogenization. A photodiode monitors the laser energy fluctuations at the fiber input. A Peripheral Component Interconnect Express (PCIE) data acquisition (DAQ) card simultaneously records the signals from the photodiode and the HNA-FUT. Wide-band preamplifiers are used before DAQ. A delay-pulse generator synchronizes the laser pulses, translation stage movement, and DAQ. Critically, this setup requires only a small amount of ultrasound gel for acoustic coupling between the imaging subject and the sample stage without water immersion. This feature preserves the normal physiological state of the subject, highlighting the practicality of the 3D-ADPAT platform for in vivo studies.

PIFER for 3D-ADPAT

Developing a tailored reconstruction algorithm is crucial for harnessing the full imaging potential of the HNA-FUT. Standard methods like the universal back-projection (UBP) algorithm are predicated on an ideal point-detector assumption, a condition that the finite-aperture effect of real-world transducer fundamentally violates. This discrepancy renders the direct application of UBP suboptimal for 3D-ADPAT. To elucidate the signal reception characteristics of HNA-FUT, we analyzed a simplified model with a point source emitting a spherical PA wavefront as the origin and the HNA-FUT surface center at a position (Δx, Δz) (Fig. 2A). Detailed calculations and analyses of the acoustic field characteristics of the HNA-FUT are provided in Materials and Methods (fig. S1). Based on the PA mechanism, a nanosecond laser pulse generates a bipolar impulse from the source, which can be modeled as a propagating narrow pressure crest and trough. The piezoelectric signal from the PZT transducer is determined by the integral of the acoustic pressure over the entire transducer surface at any given time T. The interaction between the wavefront and the transducer surface dictates the signal’s temporal profile. As the crest of the wavefront makes initial contact and sweeps across the transducer surface until T = t₁, the surface integral yields a positive value (Fig. 2B). However, because the integration angle between the transducer surface normal and the wavefront element normal is large in this intersecting region (angle α in fig. S2), only a small positive voltage is generated (Fig. 2C). As the wavefront advances, a state is reached (e.g., T = t₂) where both the crest and trough reside within the integration area. Here, their opposing pressures largely cancel, resulting in a near-zero output signal. As the wavefront proceeds further and the crest moves beyond the integration area (e.g., T = t₃), only the trough contributes to the integral. In this final stage, the angle between the normals is small (angle β in fig. S2), causing the transducer to output a large negative voltage (Fig. 2C). This process leads to a distinct signal output of HNA-FUT, which requires phase inversion to produce the positive peak for back-projection (Fig. 2D).

Fig. 2. — (A) Simulation parameter setup. TD, transducer; TOF, time of flight. (B) State of the PA point source wavefront passing over the HNA-FUT surface at T = t₁, t₂, and t₃. (C) Schematic of the output signals from an ideal transducer and the HNA-FUT, with corresponding times t₁, t₂, and t₃ marked. (D) The phase-inverted HNA-FUT signal, with TOF₁, TOF₂, and their respective deviations from the signal peak indicated. (E) Comparison of output signals from an ideal transducer (in gray) and different types of transducers (in orange). Note that the ideal signal is self-normalized, while the other detectors are normalized to the HNA-FUT’s output amplitude. CNA-FUT, conventional-NA focused ultrasonic transducer; PT, planar transducer; CQPT, conventional quasi-point transducer. (F) Dual Y-axis plot showing the changes in signal amplitude, peak-to-TOF₁ deviation, and peak-to-TOF₂ deviation for different transducers at different Δx. The amplitude for each group is normalized to its corresponding maximum value. (G) Comparison of the results of a numerical leaf vein phantom simulation for 3D-ADPAT and other imaging methods. Simulation parameters are identical except for the transducer type and reconstruction algorithm. (H) Comparison of the profiles at the same position in the numerical phantom for 3D-ADPAT and ideal transducer imaging results.

We further characterized the time-of-flight (TOF) properties of the HNA-FUT. We defined TOF₁ as the path from the point source to the transducer center, and, TOF₂ as the time of flight for the signal traveling from the point source, passing through the focal point, and traversing the focal radius f to reach the center of the transducer surface. Using the k-Wave toolbox, we simulated the signals received by various transducers (all with a 3.5-MHz center frequency and 80% bandwidth) from the same point source. The results show that the peak of the ideal transducer’s signal aligns with TOF₁, whereas the peak of the phase-inverted HNA-FUT (D = 14 mm, f = 15 mm) signal consistently aligns with TOF₂. In contrast, the peak from a commonly used PT (D = 5 mm) aligns with neither, precluding accurate reconstruction (Fig. 2E). To assess the sensitivity distribution and TOF characteristics more comprehensively, we varied the transducer’s lateral position (Δx). In terms of acceptance angle, both the HNA-FUT and a conventional quasi-point transducer (CQPT; D = 1 mm) performed excellently, maintaining more than 20% of their maximum sensitivity even at a large displacement of Δx = 46 mm. The CQPT’s low absolute amplitude, however, makes it unsuitable for high-CNR deep-tissue imaging. The acceptance apertures of the other transducers were significantly weaker (Fig. 2F). Regarding TOF behavior, the HNA-FUT signal maintained high fidelity to TOF₂ across all positions, whereas the CQPT signal was consistent with TOF₁. The other two transducers showed increasing deviation from both TOF₁ and TOF₂ with larger Δx (Fig. 2F). This analysis comprehensively demonstrates that the HNA-FUT can be rigorously reconstructed using TOF₂, thereby preserving both a large acceptance angle and high sensitivity. Furthermore, a frequency-domain comparison reveals that the surface integration effect introduces a spectral tilt toward lower frequencies for HNA-FUT, necessitating a Hamming-windowed ramp filter for spectral correction before reconstruction (fig. S3). The Hamming-windowed ramp filter (central frequency of 3.5 MHz, upper cutoff frequency of 8.5 MHz) is applied in data preprocessing, achieving the optimal trade-off between spatial resolution and ringing artifact suppression (see Materials and Methods and fig. S5).

Based on this detailed analysis, we developed the PIFER method as an integral component of 3D-ADPAT. The PIFER workflow initiates with phase inversion and filtering correction of the raw PA signals received by the HNA-FUT in the far-field, implemented using a ramp filter with a Hamming window. Subsequently, it uses a 3D back-projection where the TOF is determined by summing the distance from the field point to the focal point and the focal length (see Materials and Methods and fig. S4 for details).

We validated the approach by simulating a numerical leaf-vein phantom, using single-grid-point veins to assess the spatial resolution. Compared with the imaging results from ideal point transducers, the 3D-ADPAT image exhibits improved CNR and fidelity, with the vascular network resolved in sharp detail, albeit with a marginal decrease in resolution from 310 to 400 μm (Fig. 2, G and H). To test its robustness, we added Gaussian white noise to both raw data until the CNR of the ideal reconstruction was degraded to ~1. Under these conditions, the 3D-ADPAT image quality remained largely unaffected, confirming its resilience to noise (Fig. 2G). In contrast, the images from the PT and CQPT showed various degrees of degradation, presenting a visually evident gap in quality compared with the 3D-ADPAT result. This performance loss stems from a combination of the transducers’ inherent limitations in SNR, acceptance aperture, and TOF inaccuracy (Fig. 2G). In summary, our rigorous numerical simulations validate the powerful synergy between the HNA-FUT hardware and the PIFER method, confirming that 3D-ADPAT achieves high-CNR, high-resolution, and high-fidelity 3D PA imaging.

Performance evaluation and comparative analysis of 3D-ADPAT

Having demonstrated the advantages in theoretical analysis and simulations, we proceeded to validate the performance of 3D-ADPAT through practical imaging experiments. For all tests, the transducer was scanned with a step size of 500 μm, approximately the central wavelength, with the total number of scan points determined by the object’s dimensions. To ensure a fair comparison, identical scanning strategies and system configurations were used for all imaging methods unless otherwise specified.

First, we assessed the system’s spatial resolution by imaging a 10-μm-diameter tungsten wire with 1064-nm laser excitation and compared the results generated by PACT using a CQPT with a 1.2-mm diameter and the same central frequency and bandwidth (Fig. 3, A and B). Due to the weak signal from the thin wire, the CNR of the CQPT image was ~21. In contrast, 3D-ADPAT achieved a CNR of about 191, representing a nearly 10-fold improvement. A comparison of cross-sectional profiles reveals that the axial resolution (along the Z axis) is comparable for both methods (~170 μm for 3D-ADPAT and ~180 μm for PACT with CQPT), as axial resolution is primarily determined by the transducer’s central frequency and is largely unaffected by the synthetic aperture. However, in the XY plane, the lateral resolution of 3D-ADPAT (~420 μm, consistent with simulation results) is 2.21 times better than that generated by CQPT (~930 μm). This notable enhancement is a direct result of the large acceptance aperture of the HNA-FUT combined with the proper PIFER reconstruction. It is worth noting that the spatial resolution in 3D-ADPAT exhibits anisotropy, with the lateral resolution being weaker than the axial resolution, a consequence of the limited-view problem inherent in this scanning-based synthetic aperture configuration. A comprehensive solution to this issue will be presented in the last part of Results.

We further imaged a leaf-vein phantom using 532-nm laser excitation to compare the imaging fidelity of different methods (Fig. 3C). The 3D-ADPAT image displays a wealth of fine details with high fidelity (movie S1). In contrast, PACT with CQPT only resolves the main outline of the leaf, failing to capture the intricate internal vascular structures (movie S2). We also included a comparison with a conventional-NA focused ultrasonic transducer (CNA-FUT; D = 8 mm, f = 15 mm). Although its result shows more features than CQPT, a significant gap in quality compared with that of 3D-ADPAT remains (movie S3), underscoring the necessity of using a large-size, high-NA transducer.

To characterize the deep-tissue imaging capability of 3D-ADPAT, we fabricated a 60-mm-deep agar phantom embedded with absorbing microspheres (average diameter of ~100 μm, density of ~10 spheres/mm³). The 1064-nm laser was introduced from an XY face of the phantom. The imaging results show that 3D-ADPAT visualizes a significantly larger number of microspheres. In the CQPT results, only the larger microspheres are visible, while those smaller than the resolution limit are averaged into the background, thereby reducing the CNR (Fig. 3D). A cross-sectional comparison provides a quantitative assessment of the deep-tissue performance. At a depth of z = 5 mm, the CNR of the 3D-ADPAT image is ~58, compared to ~17 for CQPT. At z = 55 mm, the CNR for 3D-ADPAT decreases to 31, whereas the features in the CQPT reconstruction are submerged in noise, with a CNR of less than 1 (Fig. 3E). A comparison of the microsphere profiles in the 3D-ADPAT image shows that the FWHM increases by ~40% as the depth increases from 5 to 55 mm (Fig. 3F). These results demonstrate that 3D-ADPAT maintains high CNR and spatial resolution even at depths of several centimeters, significantly outperforming conventional methods.

To rigorously validate the rational system design and improved imaging capabilities of 3D-ADPAT, we conducted comprehensive comparative experiments against a near-field focal scanning mode (3D-FS), where the object is placed within the focal zone and the transducer operates strictly in focused mode (Fig. 4A). For a fair comparison, given the shallow acoustic depth of field of the HNA-FUT, we replaced it in the 3D-FS setup with the CNA-FUT yielding a larger focal zone (D = 8 mm, f = 15 mm, FZ ~ 12 mm; see Materials and Methods for details) while keeping all other acoustic parameters identical. Comparative imaging of hair (diameter of ~50 μm), microsphere (diameter of ~40 μm), and leaf vein phantoms revealed fundamental differences. Raw signal analysis indicates that 3D-FS signals are stronger but spatially localized in the focal zone and temporally narrow, whereas 3D-ADPAT signals, despite lower instantaneous amplitudes (~0.6 fold), exhibit broad temporal profiles due to the wide-angle reception of far-field wavefronts, the physical basis for high-quality PACT (Fig. 4B). While 3D-FS images rely on direct A-line stacking, 3D-ADPAT uses PIFER. This results in a critical CNR advantage: While 3D-FS CNR is limited to raw signal’s SNR, 3D-ADPAT benefits from the coherent integration of back-projected signals, acting as an effective averaging process. For instance, with ~1000 signals contributing to a single voxel (enabled by the HNA-FUT’s large acceptance angle), the CNR theoretically improves by a factor of $\sqrt{1000}$ (~31 fold). Consequently, in hair and microsphere phantoms (Fig. 4, C and D), 3D-FS suffers from severe “defocusing” artifacts and rapid resolution degradation outside the focal zone, whereas 3D-ADPAT demonstrates uniform, high-fidelity visualization of weak targets (Fig. 4, F and G). Even in the leaf vein phantom, where the 3D-FS probe was optimally positioned, it could only resolve the coarse outline (Fig. 4E), while 3D-ADPAT delineated intricate 3D vascular details (Fig. 4H). Quantitative analysis of CNR and FWHM as a function of depth (axial defocus distance) further confirms this with the depth-encoded hair images (Fig. 4, I and J). Both metrics degrade significantly faster in 3D-FS. Specifically, the CNR advantage of 3D-ADPAT expands notably with depth, exceeding sixfold at 16 mm (Fig. 4K) while maintaining a spatial resolution approximately four times better than 3D-FS at the same depth (Fig. 4L). These comparative results conclusively validate that the specialized acoustic architecture of 3D-ADPAT is capable of overcoming the depth-of-field limitations of conventional methods, enabling high-resolution, high-fidelity 3D imaging across a large, uniform FOV in deep tissue.

Fig. 4. — (A) Schematic illustrations of the 3D focal scanning mode (3D-FS) and 3D-ADPAT experimental setups, highlighting the acoustic detection regions of the transducers. (B) Raw PA signals of 3D-FS and 3D-ADPAT at the same scanning position. a.u., arbitrary units. (C to E) 3D-FS imaging results of the hair phantom, the microsphere phantom, and a whole leaf vein, respectively. MIP, maximum intensity projection. (F to H) 3D-ADPAT imaging results as comparisons. (I and J) Depth-encoded images of the hair phantom acquired by 3D-FS and 3D-ADPAT, respectively. For direct comparability, the depth origin in both images is aligned with the 3D-FS acoustic focus. Thus, the depth value represents the defocusing distance for 3D-FS. (K) Quantitative CNR comparison of the two imaging modes at different depths, displayed using a dual-color dual-axis plot. The six corresponding depth positions are marked by white dashed lines in (I) and (J). The ratio is defined as 3D-ADPAT divided by 3D-FS. (L) Quantitative FWHM comparison.

In vivo noninvasive anatomical and functional imaging with 3D-ADPAT

The performance of 3D-ADPAT was further validated through in vivo imaging. The convenient design of our platform, combined with light anesthesia, allows for flexible positioning of the animal on the sample stage. The nonimmersion design effectively prevents animal hypothermia. By coregistering the two lasers and introducing a 30-μs trigger delay between them, we can acquire dual-wavelength data nearly simultaneously while avoiding signal cross-talk. In this setup, 532 nm provides high contrast for superficial vasculature, whereas 1064 nm is used for imaging deeper organs. To mitigate motion artifacts from respiration during scanning, a gating protocol was applied to the raw data before image reconstruction. For this purpose, we used the PA signal of the sample stage to position respiratory data (Fig. 5A). This signal is suitable for gating as it does not overlap with the object’s signal and exhibits high consistency across scanning positions, allowing respiration-induced displacements to be readily detected using a cross-correlation method. For any given A-line in the raster-scan data, a correlation matrix was generated by calculating the pairwise cross-correlation coefficients between all acquired signals (Fig. 5B). This matrix was then averaged along one dimension to yield a correlation curve (Fig. 5C). Data from scanning positions with a correlation coefficient below 0.5 were classified as motion corrupted and excluded (Fig. 5D). This data exclusion process was applied to the data, and then PIFER reconstruction was performed. Both dual-color images (Fig. 5, E and G) and depth-encoded images (Fig. 5, F and H) are presented to provide a more intuitive understanding of the 3D results. In the side-lying posture, the thoracic and abdominal cavities, separated by the diaphragm, can be observed, and key organs are distinguishable (Fig. 5, E and F, and movie S4). In the prone posture, the abdominal aorta attached to the medial side of the spine and the iliac artery branches are visible, with a calculated CNR of ~37, demonstrating that 3D-ADPAT can achieve full-body penetration depth in small animals (Fig. 5, G and H, and movie S5). In contrast, the PACT using CQPT shows blurry superficial vessels with reduced resolution (fig. S6). Deeper organs are not visible due to CNR limitations, resulting in significantly degraded image quality (fig. S6). More specifically, we experimentally evaluated the validity of the dual–speed-of-sound (SOS) reconstruction method and the high quality of the raw data. Furthermore, we established reconstruction error metrics to verify the geometric fidelity of the 3D-ADPAT reconstructed images (see Materials and Methods and figs. S7 to S9). These examples illustrate the practical importance of 3D-ADPAT as a biomedical imaging tool. We also performed imaging on a living Sepia pharaonis, a cephalopod marine organism (fig. S10), where the advantages of 3D-ADPAT remained prominent, further supporting the potential for this technology in broader applications beyond biomedical research, such as marine science.

Fig. 5. — (A to D) Raw data exclusion. (A) Extract PA signals of the sample stage within each A-line. (B) The generated correlation matrix. (C) The generated correlation line. (D) A-line data with respiratory motion exclusion (red dashed box). (E to H) In vivo noninvasive anatomical imaging using 3D-ADPAT. (E) Dual-color maximum intensity projection (MIP) images of a mouse in a side-lying posture imaged by 3D-ADPAT. Features from 532- and 1064-nm excitation are represented by different colormaps. (F) The corresponding depth-encoded images in a side-lying posture. [(G) and (H)] MIP and depth-encoded images of a mouse in a prone posture by 3D-ADPAT. (I to N) In vivo noninvasive functional imaging using 3D-ADPAT. (I) Baseline images acquired before A1094 injection. [(J) to (L)] Biodistribution of A1094 at 0.5, 4, and 8 hours postinjection. Baseline anatomical structures are displayed in grayscale, overlaid with pseudocolor representing the voxel-wise percentage signal enhancement relative to the control. (M) Time-activity curves of A1094 accumulation in different liver lobes. (N) Time-activity curves of A1094 accumulation in the gastrointestinal (GI) system. HT, heart; LN, lung; DP, diaphragm; IBAT, interscapular brown adipose tissue; LV, liver; LML, left medial lobe; RML, right medial lobe; LLL, left lateral lobe; SP, spleen; ST, stomach; IN, intestine; CE, cecum; KN, kidney; AA, abdominal aorta; IA, iliac artery; TT, testicle; h, hours.

We further demonstrated the functional imaging capabilities of 3D-ADPAT by monitoring the abdominal metabolic dynamics of a specific near-infrared small-molecule probe, A1094. A1094 is a stable and rapidly cleared Second Near-Infrared (NIR-II) imaging agent that shows notable potential for clinical translation due to its sufficient blood solubility, interference resistance, and stability against pH, oxidation, and metabolism (36). This study comprehensively elucidates the probe’s spatiotemporal biodistribution and clearance pathways, providing critical pharmacokinetic data essential for assessing its metabolic fate and biosafety (8). Crucially, its absorption peak aligns with the fundamental wavelength of the Nd:YAG laser (1064 nm), making it suitable for deep-tissue imaging; thus, 1064 nm was selected as the excitation wavelength for this application. Because capturing the metabolic dynamics imposes requirements on imaging speed, we investigated the trade-offs between spatiotemporal resolution and CNR using sparse scanning strategies (fig. S11). We obtained an original reference image using a raster scan with a 0.5-mm step size over a 51 mm–by–51 mm region, corresponding to 102 by 102 sampling points (equivalent to a 10,404-element transducer array) and compared it against reconstructions with 2× to 6× downsampling factors. We observed that for large FOV (30 mm by 30 mm), the system retained volumetric imaging capabilities even at 6× downsampling (~289 effective elements) without marked quality degradation, a robustness likely attributable to the large receiving aperture of the HNA-FUT (fig. S11A). In localized views (2 mm by 2 mm), whereas strong features (CNR > 50) retained morphological integrity with only CNR attenuation (fig. S11B), weak features (CNR < 5) suffered significantly as their signals dropped below the noise floor, compounded by potential artifact interference (fig. S11C). Quantitative analysis confirmed that strong features maintained >90% correlation with the original image even at 6× downsampling (fig. S11D), whereas the correlation for weak features decayed rapidly (fig. S11E). Consequently, we adopted a 2× downsampling strategy (~2601 effective elements) for probe tracking experiment. This configuration minimizes the loss of details while reducing the DAQ time to ~4 min, which is sufficient to observe macroscopic dynamics across multiple organs.

In the experiment, we first acquired baseline images of the mouse abdomen before probe injection, revealing clear radial vascular structures within the liver (Fig. 5I). Subsequently, the A1094 solution was administered via tail vein injection, and continuous monitoring was performed for 8 hours at 30-min intervals (see Materials and Methods for details). For each time point, the amplitude change of voxels relative to the baseline was calculated and visualized as a pseudocolor overlay on the anatomical grayscale background (movie S6). Representative whole-abdominal distribution maps at T = 0.5, 4, and 8 hours are presented (Fig. 5, J to L). Overall, A1094 accumulation was primarily observed in the liver and gastrointestinal (GI) system, indicating metabolic clearance. A region of interest measuring ~0.5 mm by 0.5 mm was selected within the organ to calculate the mean accumulation change at each time point, which was then plotted as a smoothed curve. Quantitative analysis of signal kinetics across different liver lobes confirmed that the liver, as the primary metabolic organ, rapidly accumulated A1094, and interlobar heterogeneity was observed (Fig. 5M). The left and right medial lobes exhibited the most rapid accumulation, followed by a linear decay starting ~1 hour postinjection. In contrast, the left lateral lobe showed a delayed increase, peaking at ~2.5 hours, and then declining until ~4 hours before entering a plateau phase of minimal variation. This likely reflects intrahepatic hemodynamic heterogeneity: As different lobes are perfused by distinct branches of the portal vein, regional variations in perfusion efficiency and biliary excretion rates exist (37, 38). The direct vascular connectivity of the medial lobes to the main portal flow likely facilitates faster probe accumulation and clearance. In the GI tract, A1094 accumulation in the stomach followed a slow, quasi-linear increase, reaching a steady state after 5 hours (Fig. 5N). Accumulation in the intestine exhibited a sigmoidal curve, with the steepest rise occurring 4 hours postinjection (39). The cecum showed an initial increase followed by a decline, with the probe being almost cleared by T = 8 hours. These dynamics are consistent with typical hepatobiliary excretion kinetics (40): Following uptake and processing by hepatocytes, the probe is secreted into the bile, emptied into the small intestine around the 4-hour mark, transported to the cecum via peristalsis, and ultimately eliminated via the fecal route. Movie S6 provides a direct visualization of the dynamic tracking capabilities of 3D-ADPAT. Evidently, the accelerated imaging speed of 3D-ADPAT, enabled by downsampling, facilitates dynamic tracking with a raster-scanning architecture. This capability is intrinsically linked to the system’s acoustic NA and synthetic aperture reconstruction algorithms. In comparison, performing the same experiment using 3D-FS with only 2601 A-lines would require extensive interpolation for the missing data. Furthermore, restricted by the limited depth of field, 3D-FS would struggle to yield satisfactory results under these conditions. In conclusion, tracking the A1094 metabolism complements our previous anatomical findings by demonstrating the functional imaging capabilities of 3D-ADPAT. This work not only expands the scope of the manuscript beyond anatomical imaging but also robustly validates the practicality and stability of 3D-ADPAT for biomedical applications.

Array-based implementation of 3D-ADPAT

While the current 3D-ADPAT system achieves high-quality imaging through scanning, this process inherently limits its ability to capture rapid dynamic physiological events. This limitation, however, is not fundamental to the 3D-ADPAT concept and can be overcome by an array-based design, paving the way for significantly enhanced spatiotemporal resolution. Here, we present a feasible optimization scheme based on numerical simulations. A barrel-shaped array framework was designed with a bottom radius of R = 90 mm and a height of H = 100 mm. This framework houses N = 640 independently fabricated HNA-FUT elements, each with a diameter D = 8 mm and a focal length f = 9 mm (Fig. 6A). The coordinates (x_i, y_i, z_i) of the ith element are arranged on a Fibonacci grid (41) to ensure uniform distribution on the curved surface, thereby minimizing 3D imaging anisotropy

{\begin{matrix} z_{i} = (i - 1) \frac{H}{N} \\ x_{i} = Rcos [(\sqrt{5} - 1) π i] \\ y_{i} = Rsin [(\sqrt{5} - 1) π i] \end{matrix}

(3)

Fig. 6. — (A) Design of the barrel-shaped array based on a Fibonacci grid and the numerical phantom used for simulation. (B) Synthesis of an equivalent high-density array by rotating the initial sparse array through small angles. (C) Comparison of single-pulse 3D imaging (corresponding to 640 elements) and the result from the synthesized high-density array (corresponding to 6400 elements). (D) Comparison of the point spread functions (PSFs) using ideal point, HNA-FUT, and PTs. XY and YZ projections are displayed to characterize axial and lateral resolutions, respectively. Each row corresponds to the positions (Pos 1 to Pos 3) within the FOV as marked in (A). (E to G) Comparison of PSF profiles along the Y axis at positions Pos 1 to Pos 3 between the ideal point detection array [orange solid lines, corresponding to the first row in (D)] and the HNA-FUT array (blue solid lines). The axial resolutions, calculated using FWHM, are labeled. (H to J) Corresponding comparison of PSF profiles along the Z axis to quantify lateral resolution. (K to M) Comparison of imaging results in XZ projections using different element types: ideal point transducer (E), HNA-FUT (F), and PT (G). All images are depth-encoded and normalized to their respective maximum values. (N to P) imaging results in XY projections. Considering the inherently low SNR of raw data from ideal point detectors, the simulations were performed without noise to enable a more intuitive comparison of the resolution characteristics.

The simulation was configured with these parameters, and the transducer frequency response was kept consistent with the experimental scanning system. The barrel-shaped array provides panoramic PA detection of the object in the XY plane. This arrangement offers two distinct advantages: First, the uniform distribution enables isotropic 3D imaging capabilities within a specific FOV, where the frame rate is determined solely by the laser repetition rate; second, to expand the FOV, the nonrotational symmetry of the Fibonacci grid allows for image synthesis via minimal mechanical rotation, thereby minimizing the temporal overhead. In this simulation, we used a strategy of rotating the array around its central axis to sample 10 equally spaced angular positions, synthesizing a high-density aperture with 6400 virtual HNA-FUT elements (Fig. 6B). Notably, a rotation span of 8.2° was found to achieve an approximately uniform element distribution, effectively mitigating undersampling artifacts and yielding substantially improved image quality (Fig. 6C).

To quantitatively characterize the spatial resolution within the FOV, we simulated the point spread functions (PSFs) of the array-based 3D-ADPAT using isolated voxels positioned at distinct locations as sound sources. These simulations used the 10-step rotation synthesis method described above. A Cartesian coordinate system was established at the volumetric center of the array to evaluate PSFs at three specific positions: the FOV center, and 20-mm and 40-mm off-center along the Y axis (corresponding to Pos 1 to Pos 3 in Fig. 6A). To provide a rigorous benchmark, we compared the proposed 3D-ADPAT concept (reconstructed via the PIFER algorithm) against two control groups: an array composed of ideal point transducers (representing the theoretical resolution limit) and an array of conventional PTs (diameter of 6 mm, commonly used in existing systems). Both control groups were reconstructed using a conventional delay-and-sum (DAS) method. While the ideal point array offers the optimal theoretical baseline, it is physically unrealizable and would suffer from poor SNR in practice; conversely, PTs are standard but typically compromise effective aperture to maintain SNR. Maximum intensity projections in the XY and YZ planes reveal distinct differences in axial and lateral resolution (Fig. 6D). The PSF morphology of the ideal point array demonstrates that the cylindrical architecture yields better axial resolution (XY plane) compared with lateral resolution (Z direction), a slight anisotropy attributed to the panoramic coverage in the XY plane versus the finite aperture along the Z axis. Notably, due to the large detection aperture, the PSF shape remains consistent across the FOV, indicating a highly uniform imaging volume. The HNA-FUT array exhibits PSF characteristics that closely mirror those of the ideal point array. In sharp contrast, the PT array performs significantly worse; the trade-off required to maintain SNR markedly limits the effective aperture, leading to severe anisotropic distortion. The resulting PSFs are significantly broadened in the XY plane and elongated in the YZ plane (Fig. 6D), characteristics that are detrimental to high-quality 3D imaging. We quantified the resolution using the FWHM of the PSF profiles (excluding the planar array due to its visibly inferior performance). For axial resolution (profiles along the Y axis), the ideal point array yielded values of 248, 231, and 237 μm at the three positions, whereas the HNA-FUT array yielded 368, 363, and 379 μm (Fig. 6, E to G). For lateral resolution (profiles along the Z axis), the ideal point array measured 528, 547, and 568 μm, compared to 792, 826, and 786 μm for the HNA-FUT array (Fig. 6, H to J). Although the 3D-ADPAT resolution degrades by an average factor of ~1.54 (axial) and ~1.46 (lateral) relative to the ideal point limit, this performance remains exceptional for a practical system. Furthermore, we evaluated the effect of sparse sampling by comparing the rotation-free HNA-FUT array (640 elements) against the synthesized full array. Results indicate that spatial resolution does not degrade significantly with reduced element count (fig. S12, A and B); specifically, axial resolution worsened marginally from 378 to 390 μm (a factor of ~1.04) (fig. S12C) and lateral resolution from 786 to 820 μm (a factor of ~1.04) (fig. S12D). However, due to the deviation from the spatial Nyquist sampling criterion, sidelobe artifacts became more pronounced. Quantified by the SD of the background, these artifacts worsened by a factor of ~4.45 (0.0147 versus 0.0033) (fig. S12E). This quantitative PSF analysis demonstrates that the array-based HNA-FUT achieves spatial resolution approaching that of ideal point detectors, suggesting that the large NA effectively mitigates the dependency on element density.

To validate these findings in complex biological structures, we further simulated the imaging performance of the array-based 3D-ADPAT on a leaf vein phantom, benchmarking it against arrays composed of ideal point transducers and conventional PTs (diameter of 6 mm). Consistent with the quantitative PSF analysis, the 3D-ADPAT reconstruction (via PIFER) (Fig. 6, K and N) demonstrates imaging clarity only marginally inferior to that of the ideal point array (Fig. 6, L and O) while crucially retaining the advantage of substantially higher sensitivity required for deep-tissue imaging. In sharp contrast, the PT array suffers from severe image degradation, characterized by significant lateral blurring (Fig. 6M) and a radial decay of axial resolution away from the FOV center (Fig. 6P). This comparison confirms that the degradation observed in conventional planar arrays is a direct consequence of the finite-aperture effect (limited acceptance angle), a fundamental physical constraint that cannot be rectified merely by increasing element density.

To further validate the feasibility of this approach, we have envisioned a specific system implementation strategy (fig. S13, A and B). The 640 elements can be housed within a stainless steel casing, with the imaging subject positioned along the central axis. Openings in the casing accommodate fiber bundle branches to ensure uniform optical excitation of the subject. Operationally, the system can use a common pulsed laser with a 10-Hz repetition rate, with the array mounted on a rotary stage ~30 cm in diameter. Laser triggering, mechanical rotation, and DAQ are strictly synchronized by a timing controller. In scenarios requiring high-speed imaging, the array remains stationary, achieving 10-Hz volumetric imaging within a limited FOV. Conversely, for large-FOV dynamic imaging needs, the rotary stage and laser can be triggered 10 times per second, rotating 0.82° per step for a total of 8.2°. This process synthesizes an effective aperture of 6400 detection elements within 1 s, enabling 1-Hz volumetric imaging for a large FOV. The rotary stage operates in a continuous small-angle reciprocating rotation mode, synthesizing images identically during both clockwise and counterclockwise movements. Calculations indicate a rotation speed of ~8.2°/s, which is mechanically undemanding. Such rotary stages are mature commercial products, and their rotational and repositioning accuracy (on the micron scale) meets the requirements for PACT imaging. The frame rate for large-FOV imaging can be enhanced by increasing the laser repetition rate. For instance, upgrading to a 50-Hz laser would increase the volumetric imaging frame rate to 5 Hz. Because lasers with this repetition rate are mature commercial products and high-end rotary stages are readily available, there are minimal hardware complexities associated with this upgrade. Therefore, the proposed scheme is feasible.

Upon reexamining the contrast between single-pulse and rotation-synthesized imaging, it is evident that despite the significant artifacts present in single-pulse imaging across a large FOV, the discrepancy diminishes to an insignificant level within a sufficiently small FOV (e.g., 10 mm) (fig. S13, C and D). This comparison compellingly demonstrates that 3D-ADPAT is capable of supporting high-quality, real-time 3D imaging within a restricted FOV (such as hemodynamics or neuroimaging), while the rotation synthesis strategy effectively ensures high performance across a large FOV (such as whole-body metabolic monitoring or multiorgan tracking).

Existing studies using four 256-element arc-shaped transducer arrays have demonstrated satisfactory 3D dynamic imaging quality by performing continuous high-speed reciprocating rotation over 90° (7, 42, 43). Their high-level industrial design is specifically required to ensure stability when driving large arrays through such large angles at high speeds. In comparison, due to the unique acoustic reception design of 3D-ADPAT, fewer elements are required, allowing for more compact packaging. Crucially, the substantial reduction in the required rotation angle significantly alleviates the mechanical burden. Consequently, the array-based 3D-ADPAT, based on the cylindrical Fibonacci array and continuous small-angle rotation, has substantial advantages in system robustness for achieving large-FOV, high spatiotemporal resolution dynamic 3D imaging.

DISCUSSION

In this work, we have developed and validated 3D-ADPAT, an imaging system that leverages a simple HNA-FUT in conjunction with a bespoke PIFER algorithm. This synergistic hardware-software approach enables deep-tissue imaging with high CNR and high spatial resolution, performances of which have been rigorously validated through numerical simulations, phantom experiments, and diverse in vivo examples. Our simulations confirmed that 3D-ADPAT can approximate the 3D imaging quality of an ideal point detector while exhibiting strong robustness against noise. Experimental comparisons with conventional methods using a quasi-point PZT transducer demonstrated that 3D-ADPAT achieves a nearly 10-fold improvement in CNR and a 2.21-fold enhancement in lateral resolution, facilitating high-quality visualization of targets at depths up to 6 cm in tissue-mimicking phantoms and across the entire torso of small animals. The use of the HNA-FUT is fundamental to this success as it provides a large reception aperture without requiring external diffractive elements (e.g., slits or pinholes) and, crucially, without sacrificing the high SNR essential for achieving both high resolution and deep penetration. The PIFER algorithm, refined through meticulous simulations, computationally unlocks the full potential of the rich information captured by the HNA-FUT. Here, we emphasize that 3D-ADPAT is a PACT mode, where image reconstruction relies on the back-projection of signals from the entire scanning geometry, a synthetic aperture process, meaning that every transducer element contributes to the reconstruction of every voxel within the FOV. This differs from PA microscopy (PAM), where images are formed by directly mapping A-line data from the focal zone, with each transducer contributing only to the voxel at the current scan position (44, 45). Moreover, the inherent trade-off between the focal zone length (depth of field) and the focal spot size poses a significant challenge for PAM in achieving high-resolution deep-tissue scanning (46).

A pivotal advantage of 3D-ADPAT is its inherent scalability as the HNA-FUT is based on conventional piezoelectric materials that are amenable to cost-effective mass production. This feasibility paves the way for array-based implementations. We have proposed and simulated a barrel-shaped array configuration where HNA-FUT elements are distributed on a Fibonacci grid. By synthesizing a high-density array through small-angle mechanical rotation, this design can theoretically achieve dynamic 3D imaging with high CNR, near-isotropic spatial resolution, a centimeter-scale large FOV, and a hertz-level temporal resolution. Compared with conventional curved arrays constructed from planar elements, the 3D-ADPAT barrel array offers significant improvements in both CNR and lateral resolution, a conclusion consistent with our findings from the scanning-based system. This panoramic detection architecture not only mitigates the anisotropy between axial and lateral resolution but also enables true global imaging of living animals, rather than being restricted to a single ventral or dorsal view (47, 48). Furthermore, the single-pulse 3D imaging capability of the array-based 3D-ADPAT can facilitate rapid imaging within a subcentimeter FOV, with the frame rate being limited only by the laser repetition rate (e.g., 20 Hz). Therefore, the scalable FOV and spatiotemporal resolution of the array-based 3D-ADPAT, combined with its potential for multispectral molecular imaging, present an effective strategy for the development of 3D PA imaging platforms. Notably, implementing such an array configuration using 3D-FS approach would be prohibitively difficult. The 3D-FS paradigm, which relies on focused detection and A-line stacking, struggles with array integration; specifically, stacking A-lines in 3D space necessitates extensive nonuniform interpolation, leading to severe FOV distortion, a limitation frequently observed in photoacoustic/ultrasound endoscopy (49, 50). Accordingly, the combination of synthetic aperture reconstruction and optimized array design highlights the distinct advantage of the 3D-ADPAT architecture.

To intuitively highlight the advantages of 3D-ADPAT, we have summarized relevant technologies into a comparative table (table S1). Early attempts in this direction involved attaching custom-made acoustic lenses to flat transducers to expand the acoustic aperture (28). Although these pioneering studies were relatively rudimentary regarding reconstruction algorithms and performance validation, they provided critical inspiration for subsequent research. Subsequently, cylindrical focused transducers (typically commercial probes) were used for rotational 2D imaging. However, because they have strong focusing in only one direction while the aperture in the orthogonal direction remains restricted, the improvement in overall imaging quality is inherently limited (26, 27). With the development of the field, center-focused ring arrays have been widely adopted for 2D tomography. While the elongated elements of ring arrays provide high SNR, the 3D images formed by stacking longitudinal scans suffer from significant resolution anisotropy (9). Although synthetic aperture focusing techniques can mitigate this issue to some extent, the improvement is limited. Fundamentally, because the imaging target is located at the array’s geometric center where all element foci converge, it resides within a “near-field” region. Consequently, this configuration fails to effectively use the far-field aperture required for high-quality isotropic 3D imaging. 3D imaging based on commercial linear arrays (e.g., the common Philips L7-4) has gained attention due to its high flexibility. When combined with focal-line reconstruction algorithms, the lateral resolution is improved compared with that of the 2D stacking results (51, 52). However, two persistent issues remain: First, commercial linear arrays typically use elements with conventional or weak elevation focusing to maintain slicing capability, which inherently restricts their far-field acoustic aperture; second, the calculation of TOF based on the focal-line assumption often lacks rigorous simulation analysis, raising questions about the accuracy of the reconstruction parameters. By installing a slit at the focal-line position, the diffracted wavefront becomes conjugate to the transducer surface, effectively expanding the NA (31–33). In 3D images reconstructed via scanning, the lateral resolution is noticeably improved. Over years of development, researchers have used this structure to achieve whole-body structural-functional imaging in small animals (34). Nevertheless, the drawbacks of this approach are pronounced: First, the introduction of the slit markedly reduces the SNR of the raw data, making deep-tissue imaging significantly challenging; second, the reliance on external precision mechanical attachments often hinders flexible system integration and scalability. In contrast, our 3D-ADPAT system uses a spherically strongly-focused HNA-FUT, physically achieving a large receiving aperture without compromising sensitivity. Systematic simulations of TOF have validated the rationale of our proposed PIFER method. Furthermore, phantom and in vivo structural-functional imaging experiments have confirmed its high spatial resolution and fidelity, and simulations have outlined a viable pathway for array-based systemic integration to achieve high spatiotemporal resolution. Consequently, we anticipate that 3D-ADPAT holds potential to advance the clinical translation of PA modalities, particularly for scenarios requiring deep-tissue imaging, such as whole-breast imaging.

Despite the advantages conferred by the HNA-FUT, its extended surface area introduces physical phenomena that warrant discussion. Subtle “ghosting” artifacts, appearing as a faint shell around reconstructed objects, can be observed in some 3D-ADPAT results (Fig. 6O). This is a manifestation of the first-arrival wave, an effect inherent to the surface integration process. Although the amplitude of this first-arrival signal is small compared with the primary signal used for reconstruction, its impact on the quantitative accuracy of PA images requires further evaluation. Deconvolution-based methods may offer a viable strategy to mitigate this artifact (53, 54). To evaluate the potential improvements in image quality using deconvolution, we performed assessments with both phantom and in vivo data. Initially, we experimentally characterized the 3D PSF of the system by imaging 20-μm nickel-coated polystyrene microspheres under 532-nm excitation (fig. S14A). To ensure a high SNR, the laser pulse energy was maximized to ~400 mJ per pulse. Using this PSF, we applied the Richardson-Lucy (R-L) iterative deconvolution algorithm (implemented in MATLAB) to the hair phantom data (55, 56). Visual inspection revealed that the hair features in both XY and YZ projections became notably sharper after four deconvolution iterations (fig. S14B). With increasing iterations, the features in zoomed-in views exhibited enhanced sharpness in both lateral and axial dimensions, accompanied by a visible suppression of the “first-arrival” artifacts (fig. S14C). For quantitative analysis, we extracted lateral and axial profiles to measure the FWHM to characterize spatial resolution and the signal-to-artifact ratio, defined as the ratio of the feature peak to the peak of the first-arrival artifact, to assess artifact severity. Specifically, in the XY projection, after eight iterations, the lateral feature width improved from an initial 805 to 426 μm, while the ratio increased from 5.24 to 10.75 (fig. S14D). In the YZ projection, the axial feature width was optimized from 199 to 93 μm, with the ratio improving from 5.66 to 8.66 (fig. S14E). These metrics quantitatively demonstrate the feasibility of using deconvolution to achieve super-resolution and suppress first-arrival artifacts under ideal conditions. However, the method’s performance on in vivo data was less robust. Increasing the iteration count led to the fragmentation of continuous features (fig. S14F), although vascular structures became clearer in certain local regions (fig. S14G). We attribute this limitation primarily to two factors: (i) Statistical model mismatch, as the R-L algorithm is derived based on Poisson statistics (typical for optical photon counting), which is likely inconsistent with the noise distribution inherent to PACT (57, 58); (ii) PSF space variance, as the standard R-L implementation assumes a spatially invariant PSF, whereas, in biological tissue, the PSF degrades and varies with depth due to tissue heterogeneity and frequency-dependent attenuation. On the basis of these findings, we conclude that, while deconvolution is effective for resolution enhancement and artifact suppression in controlled settings, its generalization to in vivo scenarios is now limited. Consequently, future implementations of 3D-ADPAT would require deconvolution strategies specifically optimized for these challenges to exploit this potential.

Penetration depth remains a cardinal challenge in PACT, distinguishing it from purely optical modalities. While 3D-ADPAT uses the HNA-FUT to maximize detection aperture and secure imaging depth, further extending this limit necessitates advanced hardware strategies. Although acoustic lenses could theoretically expand the aperture and enhance coherent summation, their reliance on temperature-sensitive refraction introduces focal instability, posing challenges for the precise focal parameters required by PIFER reconstruction. A more robust alternative for achieving large-aperture spherical focusing lies in the “dice-and-splice” piezoelectric composite technique (59), which avoids these thermal drift issues. Ultimately, maximizing penetration depth requires a holistic synergy across optics, electronics, and algorithms. Optically, effective photon management via optimized illumination geometries, such as dark-field or multifiber patterns, is critical for delivering photons to deep targets (60, 61). Electronically, beyond impedance matching, we emphasize the often-underestimated role of dynamic range in DAQ. Upgrading from standard 12-bit to 16-bit systems is essential to resolve weak deep-tissue signals that would otherwise fall below the quantization noise floor, rendering them unrecoverable even by synthetic aperture averaging. Furthermore, to address in vivo acoustic heterogeneity, we anticipate that the rapid growth in computing power will facilitate the adoption of 3D full-waveform inversion (3D-FWI) (62). By correcting phase aberrations through high-resolution speed-of-sound mapping, 3D-FWI holds the potential to unlock high-fidelity, ultradeep PACT.

While PACT is unique in its capacity to resolve deep-tissue spectral information, several critical challenges must be addressed to unlock its full potential. First, the penetration depth is now constrained by the performance limits of optical, acoustic, and electronic components, as well as existing reconstruction algorithms, which restricts many deep-organ clinical applications. Second, quantitative imaging remains difficult due to the in vivo “spectral coloring” effect: The unknown fluence attenuation complicates spectral unmixing, making it challenging to derive precise biological biomarkers from deep tissues (63). Third, the high dependence on bulky hardware, such as large transducer arrays and high-energy lasers, limits system integration, preventing the transition toward miniaturized or wearable imagers. Future advancements in highly integrated laser sources will be pivotal in driving this transformation. Fourth, the field lacks specialized molecular probes engineered specifically for PA characteristics, such as high-response near-infrared proteins (64). Historically, the development of specific probes (e.g., fluorescent proteins) has been a primary catalyst for expanding imaging modalities, and similar breakthroughs are essential for PACT. Despite these hurdles, the application prospects for PACT remain robust. In the near term, clinical applications focusing on accessible tissues, such as characterizing dermatological conditions (65) or monitoring peripheral vascular dynamics (e.g., in diabetic foot ulcers), offer immediate translational value (66). Looking further ahead, the integration of embodied artificial intelligence and robotics represents a paradigm shift. Unlike current bulky systems (e.g., stationary breast imagers) that require the patient to adapt to the machine, future PA robots will have the intelligence to adaptively scan and serve the patient (67). Last, the rise of ultrasound neuromodulation highlights a unique niche for PACT. By combining PACT with genetically encoded calcium or voltage indicators, it becomes possible to visualize deep-brain neural activities. A dual-modality PACT/ultrasound system could establish a “structure-modulation-function” visualization link, offering a powerful pathway for neuroscience and brain research (68, 69).

The 3D-ADPAT architecture may also be adaptable for ultrasound imaging, given the shared requirements for emission/reception aperture and SNR. However, in a pulse-echo ultrasound modality, the first-arrival wave effect would be amplified as the transducer is used for both transmission and reception. Concurrently, the ramp filter used in PIFER to correct for the spectral distortion caused by the integration effect inherently attenuates some low-frequency components not originating from this effect. While this has a minor impact on PA imaging, it could impose limitations on certain ultrasound methods that rely on frequency components in the hundred-kilohertz range, such as ultrasound FWI (70, 71). Notably, the 3D-ADPAT system shares a parallel cross-disciplinary design philosophy, specifically in maximizing wide-angle aperture and detection sensitivity, with recent optical metalens arrays (72). Guided by this shared insight, future PACT systems could adopt miniaturized ultrasonic arrays designs to achieve panoramic coverage within a compact housing, thereby facilitating the monitoring of ultrafast physiological dynamics.

The integration of deep learning also represents a transformative frontier in modern imaging, particularly in its potential to alleviate hardware dependencies through algorithmic compensation. Drawing parallels to the optical domain, where end-to-end networks have successfully corrected aberrations in compact metalens-integrated cameras (73, 74), deep learning in 2D PACT has similarly facilitated high-quality reconstruction from sparse or limited-view data (75). However, translating these successes to 3D PACT remains challenging due to the severe spatial sparsity inherent in 3D geometries, where distributing sensors over a surface results in significantly higher sparsity compared with a curve. Notably, the 3D-ADPAT addresses this by significantly increasing the information throughput of individual elements. We envision that the synergy between sparse 3D-ADPAT arrays and deep learning could pave the way for handheld, real-time, high-fidelity 3D PACT devices. Furthermore, inspired by self-super-resolution advancements in optical coherence tomography (76), we highlight the critical necessity of self-supervised learning for PACT, given the practical difficulty of obtaining in vivo ground truth (77). It is crucial to note, however, that directly transferring optical strategies (e.g., Noise2Noise) (78) is nontrivial. Unlike statistical noise distributions in optical images, PACT suffers from structure-dependent artifacts that require tailored network architectures for effective disentanglement. Consequently, future research must focus on developing self-supervised frameworks specifically designed for the unique physics of PA imaging.

In conclusion, 3D-ADPAT emerges as a highly promising and accessible imaging method that effectively overcomes long-standing challenges in PA tomography. Its outstanding performance in terms of spatial resolution, imaging depth, and CNR, coupled with its potential for high-spatiotemporal dynamic imaging, establishes it as a powerful tool for biomedical research. By providing a practical technological pathway, 3D-ADPAT offers substantial inspiration for the development of advanced PA systems.

MATERIALS AND METHODS

3D-ADPAT system setup

The HNA-FUT and all other transducers used in this study were custom fabricated from piezoelectric composite materials (ULSO TECH Co. Ltd.). The HNA-FUT features a focal length of 15 mm and a lateral diameter of 14 mm. The transducers were designed to be self-focusing, with a central frequency of ~3.5 MHz and a bandwidth of around 80%. This self-focusing design ensures that the focal position remains stable and independent of SOS in water, providing a consistent coordinate system for reconstruction. The transducer’s matching layer was fabricated in a dark gray color to generate a surface signal, which serves as a temporal fiducial for the laser firing event (t = 0). The acoustic field of the transducer was characterized and visualized using a beam analyzer (OS-12D, ONDA Co. Ltd.). The imaging sample stage was constructed from a 10-mm-thick agar plate (2.5% w/v), which provided both optical and acoustic transparency. This agar plate was embedded in a white resin frame, which could be fitted with auxiliary components for animal imaging, such as respiratory tubes. The imaging subject was placed on the upper surface of the stage, and acoustic coupling was achieved by applying a thin layer of standard ultrasound gel without water immersion. The region below the sample stage was submerged in deionized water. The transducer was mounted on a cantilever arm attached to a two-axis motorized translation stage (L-509.20sd00, 52-mm travel distance, 0.2-μm positioning accuracy, PI Inc.), enabling matrix scanning. The scanning range, step size, and speed were controlled via a custom-developed graphical user interface.

Two Nd:YAG lasers (Nimma-900, 10-Hz pulse repetition rate, Beamtech Optronics Co. Ltd.) were used as excitation sources. One laser provided output at 1064 nm, whereas the other was equipped with a second-harmonic generation module to produce output at 532 nm. The two beams were coaligned using a silver mirror and a dichroic mirror before being coupled into a bifurcated quartz fiber bundle (Nanjing Chunhui Science Technology Industrial Co. Ltd.). Each of the two output terminals of the fiber bundle was shaped into a 30 mm–by–1 mm line window and mounted on a 3D printed resin holder. An adjustable silver mirror at the end of each holder directed the diverging beam toward the imaging FOV. This symmetric, dual-sided illumination architecture ensured that the beams overlapped to create a uniform illumination field and that the light path did not interfere with the transducer’s scanning area. The weak scattering of the agar plate further homogenized the light distribution.

The signal from the transducer was first passed through a low-pass filter (BLP-15+, Mini-Circuits Co. Ltd.) to remove high-frequency electromagnetic interference, followed by two cascaded broadband preamplifiers (ZFL-500LN+, Mini-Circuits Co. Ltd.) to enhance the SNR. A PCIE DAQ card (PCIE-14100-x8, 14-bit, Beijing Yixin Technology Co. Ltd.) performed the Analog-to-Digital (A/D) conversion at a sampling rate of 300 MS/s (300 Mega-Samples per second). A photodiode (PDA10A2, Thorlabs Inc.) was placed near the fiber bundle and connected to the DAQ card to monitor laser energy fluctuations for subsequent data correction. A digital delay and pulse generator (DG645, Stanford Research Systems) synchronized the timing of the laser firing, stage movement, and DAQ.

For imaging experiments, the step size for both the fast and slow axes of the scanning stage was set to 500 μm. A full scan over the maximum travel range of both axes took ~16 min. At each scanning position, a pair of dual-wavelength laser pulses, separated by a 30-μs delay, was fired, and the corresponding PA signals (from the transducer) and light energy data (from the photodiode) were acquired as a set.

For the comparative 3D-FS experiments, the transducer was replaced with a CNA-FUT and adjusted in close proximity to the target to ensure operation in focused detection mode. All other hardware configurations and experimental protocols remained consistent.

Analysis of the acoustic field for a focused ultrasonic transducer

3D-ADPAT requires the HNA-FUT to minimize the focal zone size, approximating an ideal virtual point for far-field detection It has a diameter D = 14 mm and focal length f = 15 mm. Here, a parameter analysis is given (fig. S1). For such a strongly focused element, the NA is calculated as (79, 80)

NA = \frac{D}{\sqrt{4 D^{2} + f^{2}}}

(4)

Substituting the values yields NA = 0.423, characterizing a typical high-NA probe. The lateral focal diameter df (−6 dB) and the focal length FZ (−6 dB) are defined respectively as (81–83)

df = 1.02 \frac{λ}{NA}

(5)

FZ = \frac{2 λ}{{NA}^{2}}

(6)

With a center frequency of 3.5 MHz and speed of sound c = 1490 m/s (resulting in λ = 426 μm), we calculate df of ~1.03 mm and FZ of ~4.7 mm. Crucially, due to this strongly focused geometry, the Rayleigh distance formula (typically used for planar pistons) is inappropriate for determining the operational zone here. Instead, the transition boundary N between the Fresnel region (near field) and the Fraunhofer region (far field) for a focused transducer should be approximated by (84)

N = f + \frac{D^{2}}{8 f}

(7)

Substituting the parameters yields N = 16.63 mm. In our system implementation, the distance from the transducer surface to the bottom of the agar stage is ~77 mm, plus a 10-mm sample stage. This ensures that the entire imaging object is situated well within the Fraunhofer field of the transducer (because 77 mm > 16.63 mm), supporting our operational assumption. We also acknowledge that treating the HNA-FUT focus as a “virtual point detector” is a conceptual abstraction derived from TOF characteristics. Physically, the propagation of wavefronts is independent of the HNA-FUT’s geometric focus; the waves merely follow acoustic propagation laws and are physically integrated upon impinging on the detector surface. This abstraction provides an intuitive yet profound understanding of how the HNA-FUT achieves both large acceptance aperture and high SNR.

Numerical simulation setup

All numerical simulations presented in this study were conducted using the k-Wave toolbox, a powerful open-source acoustic modeling package for MATLAB (R2023b). This toolbox was selected for its proven accuracy in solving the coupled acoustic wave equations using a k-space pseudospectral method, which is highly efficient for large-scale, 3D simulations. The computations were executed on a high-performance workstation equipped with an Intel Xeon Silver 4210R CPU (40 cores at 2.4 GHz) and 384 GB of RAM. To significantly reduce computation time, the forward PA propagation process was accelerated using a GPU (NVIDIA Quadro RTX 8000, 48-GB VRAM, CUDA version 11.6).

In all simulations, the transducer models were configured with a center frequency of 3.5 MHz and an 80% fractional bandwidth to faithfully replicate the characteristics of the experimental hardware. A uniform spatial grid size of 200 μm and a temporal sampling rate of 14 MHz were used. These values were chosen to ensure numerical stability, satisfying the Courant-Friedrichs-Lewy condition, and to provide ideal sampling of the signal bandwidth without aliasing. The medium SOS was set to 1490 m/s. The integration effect of different transducer geometries (e.g., focused and planar) was accurately modeled by defining the transducer surface as a binary mask of active grid points and summing their individual pressure signals at each time step, a method that maintains physical consistency with the real-world detection process.

The study involved three distinct and progressively complex simulations. First, the point-source-to-single-transducer simulation (Fig. 2, A to F), designed to elucidate the fundamental signal reception physics, was conducted on a 260 by 110 by 375 grid. Each forward propagation in this setup required ~50 s. Second, the leaf-vein scanning simulation (Fig. 2, G and H), which aimed to validate the PIFER algorithm, was performed on a larger 300 by 300 by 256 grid, covering an effective scanning area of 51 mm by 51 mm. The transducer surface was positioned about 70 mm from the numerical phantom, mimicking the experimental geometry. A single forward propagation took ~2 min. To accelerate this computationally intensive simulation, we implemented a parallelization strategy by arranging the elements in a 3 by 3 array with a 17-mm pitch, generating nine channels of data in one propagation. To synthesize the final image with the desired 500-μm scanning step size, this 3 by 3 array was scanned in a zigzag pattern over a 17 mm–by–17 mm area, requiring (17 mm/500 μm)² = 1156 steps. This strategy reduced the total simulation time to ~38 hours, representing a ninefold improvement over a naive single-element scanning approach. Last, the barrel-shaped array simulation (Fig. 6), exploring a future high-speed implementation, was executed on a very large 1024 by 1024 by 640 grid. A single forward propagation took about 20 min, and the complete simulation for the 10 synthesized rotational angles required ~3.5 hours.

3D image reconstruction

All 3D PA image reconstruction algorithms were implemented in MATLAB R2023b and accelerated using GPU parallel computing on the same workstation described previously. Taking into account the transducer’s central frequency, the reconstruction grid for all images was defined with an isotropic voxel size of 0.1 mm by 0.1 mm by 0.1 mm.

The 3D-ADPAT reconstruction uses the PIFER method to achieve high-quality volumetric imaging from the HNA-FUT data. The workflow consists of the three key steps (fig. S4):

1) HNA-FUT DAQ: The process begins with raw PA signals acquired by the HNA-FUT. This specific transducer configuration provides a large receiving aperture and a high SNR, establishing the essential physical foundation for high-resolution 3D-ADPAT imaging.

2) Correction for the integration effect: To eliminate the signal integration effect resulting from the extended detection area, the raw data undergo preprocessing. This involves phase inversion followed by the application of a spectral ramp filter apodized by a Hamming window. The filter parameters are configured with a peak frequency of 3.5 MHz (matching the transducer’s center frequency) and a cutoff frequency of 8.75 MHz.

3) DAS reconstruction and directivity compensation: A 3D-DAS algorithm is subsequently applied. For each scanning position, signals are spherically back-projected into the 3D imaging space. The TOF is calculated as the sum of the distance from the field point (voxel) to the transducer’s focal point and the fixed focal length (distance from the focus to the transducer surface) of HNA-FUT. The radius of the projection sphere is determined by the sum of the products of the calculated TOF and the SOS for each segment along the line connecting the voxel to the focal point. To ensure amplitude fidelity, a cosine directivity factor is applied to each projection wavefront. Specifically, the reconstructed value for each voxel is weighted by the cosine of the angle between the vector connecting the voxel to the focal point and the transducer’s surface normal.

For the conventional focused transducers, the same PIFER method was applied; however, its smaller NA resulted in suboptimal performance. For the ideal point and quasi-point transducers, a standard DAS algorithm was used, with the TOF calculated directly from the field point to the transducer’s position. For the PT, where a physically precise TOF is difficult to obtain, the TOF was also calculated from the field point to the transducer center, followed by DAS reconstruction.

For experimentally acquired PA data, each A-line signal was first normalized by the integration value from the corresponding photodiode measurement to correct for laser energy fluctuations. In in vivo imaging, the SOS difference between the animal body and the surrounding water/agar medium can cause image blurring if a uniform SOS is applied in reconstruction. To address this, we implemented a dual-SOS reconstruction method. First, an initial, nonideal reconstruction was performed with uniform SOS to locate the animal’s lower surface. This interface was then used to segment the 3D grid into upper (animal) and lower (medium) regions, each with a distinct SOS. For any given field point, the back-projection path was calculated by first determining the coordinates of its intersection with the interface and then summing the two resulting path segments, each with its respective SOS. For 532-nm imaging, which due to its limited penetration depth primarily captures feature from superficial vessels, the PA signals propagate mainly through the lower region. Therefore, we first set the SOS for both regions to 1490 m/s and iterated the lower region’s SOS to maximize the peak CNR and sharpness of the surface vessels, resulting in an optimal SOS of 1484.6 m/s. For 1064-nm imaging, which targets deeper organs, the lower region’s SOS was fixed at the previously determined value, and the upper region’s SOS was iterated in a similar manner, resulting in an optimal value of 1495.2 m/s for the animal body.

To provide direct validation for the proposed iterative method and the rationality of our SOS values, we conducted comprehensive supplementary experiments (fig. S7). We first constructed a setup capable of accurately measuring the SOS of uniform cylindrical objects. The apparatus consists of a PT (central frequency of 3.5 MHz, diameter of 7 mm), a polished aluminum alloy acoustic reflector (10 mm thick), and an ultrasonic pulse transceiver (fig. S7A). The transceiver triggers the probe to emit transmission signals and receives the echoes, while an oscilloscope acquires the waveforms. It is worth noting that this setup is designed to test simple cylindrical objects as irregular interfaces or scattering distributions would interfere with peak detection and introduce measurement errors. The reflector was fixed to the bottom of a water tank, and the probe was mounted on a multiaxis adjustable stage. System calibration was achieved by fine-tuning the transducer orientation to maximize the echo from the reflector, ensuring the transducer face was parallel to the reflector surface. We fabricated two cylindrical agar phantoms with distinct SOS properties using high-precision 3D printed molds (height l = 28.55 mm). One was a pure-water agar phantom (2.5% w/v), and the other substituted water with a 20% (w/w) sucrose solution while maintaining other components. During testing, the probe contacted the phantoms without compression, and the position was kept consistent via the stage. By measuring the time interval Δt between the probe surface echo and the reflector echo, the SOS was calculated as v = 2 l/Δt. It yielded the SOS of 1490.9 m/s for both the water environment and the pure-water phantom, and 1552.9 m/s for the sucrose phantom (fig. S7B). These results are consistent with the established physical fact that sucrose increases the speed of sound (85), as well as the known temperature dependence of sound speed in water at room temperature (~23°C) (86). We then used a hair phantom to verify the rationality of the dual-SOS iterative method. We created a hair phantom using the aforementioned sucrose-agar solution and performed 3D-ADPAT imaging at the same ambient temperature (fig. S7C). Because the hair was encapsulated in agar, the signal traveled through two distinct SOS media: the sucrose phantom and the water environment (fig. S7D). Similar to the animal imaging, we defined a 3D ellipsoid roughly enclosing the phantom and set the internal and external sound speeds (v_in and v_out) (fig. S7C). With v_out fixed at 1490.9 m/s, we initialized v_in at 1490.9 m/s and iteratively reconstructed the image using feature’s CNR in the FOV center as the optimization parameter, with a step of 0.1 m/s (fig. S7E). The v_in-CNR curve peaked at v_in = 1550 m/s, which agrees remarkably well with our ground-truth measurement (1552.9 m/s). In the reconstruction results, when v_in = 1490.9 m/s, the hair features appeared blurred and distorted (fig. S7F), whereas, at v_in = 1550 m/s, the features became distinct (fig. S7G). Profiles at the same location further highlighted this difference. Incorrect SOS reconstruction resulted in low CNR and signal broadening, while the correct reconstruction yielded sharp features with high CNR (fig. S7H).

For enhanced visualization of vascular structures in in vivo images, a Hessian-based Frangi vesselness filter (87) was applied to the reconstructed volumes. The final images were generated by weighted fusion of the self-normalized filtered image (with a weighting factor of 0.2) with the self-normalized initial image (with a weighting factor of 0.8). These steps are standard postreconstruction visualization techniques widely adopted in medical angiography (e.g., computed tomography or magnetic resonance imaging) (87, 88) and high-impact PA studies to accentuate vascular features (89, 90). High-quality raw data, characterized by low common-mode cross-talk, serves as the foundation of our approach (fig. S8A). While Frangi filtering can be criticized for potentially enhancing noise into vessel-like artifacts, this liability is effectively mitigated in 3D-ADPAT. The high CNR of our raw reconstructions suppresses the generation of such false features. To demonstrate this visually, we have added comparative images of in vivo superficial vessels before and after Frangi filtering. It is evident that the raw images reconstructed by PIFER already have high quality with clear vascular features (fig. S8B). The Frangi filter serves merely as a visualization enhancement (fig. S8C). A detailed comparison reveals negligible false feature generation. Crucially, we emphasize that all quantitative analyses in this study (e.g., calculating percentage changes in probe accumulation) are performed strictly on the raw images without any postenhancements. In addition, quantitative reconstruction error metrics are vital for assessing imaging reliability. We evaluated the geometric accuracy by comparing the photo of the tungsten wire phantom with its 3D-ADPAT reconstruction results. In the optical photograph of the phantom, the angle between the two wires is 84.7° (fig. S9A). In the 3D-ADPAT’s result, by connecting the peak positions of the cross sections as reference lines, the calculated angle is 84.1° (fig. S9B). The absolute relative angular error is only 0.71%, demonstrating the high geometric fidelity and reliability of 3D-ADPAT.

To reconstruct the 3D-FS volumetric images, the raw A-line signals (102 by 102 grid) acquired via raster scanning were initially arranged based on the 0.5-mm scanning step. To refine visualization, bilinear interpolation was applied to upsample the lateral grid to a 0.1-mm spacing. Simultaneously, in the axial direction, the A-line signals were mapped to the spatial domain and resampled at 0.1-mm intervals based on the speed of sound. This process yielded a final 3D volume with an isotropic voxel size of 0.1 mm by 0.1 mm by 0.1 mm.

The majority of the 3D volumetric visualizations in this manuscript were rendered using the built-in Volume Viewer toolbox in MATLAB R2023b. Depth-encoded images were generated using custom MATLAB scripts. For in vivo probe tracking applications, the fusion of pseudocolor and grayscale data was performed using Amira 2019 software.

Selection of ramp filter windows

In 3D-ADPAT, the HNA-FUT exhibits a unipolar-like waveform at large oblique incidence angles (Fig. 2). This phenomenon is attributed to the spatial integration effect of the expanded aperture. Because a standard ramp filter inherently amplifies high-frequency noise (fig. S5A), the proposed PIFER method uses a Hamming-windowed ramp filter for data processing.

For validation, we explored the effects of five different windowing schemes on filtering (91–93) by imaging the hair phantom (with the ramp filter’s upper cutoff frequency set to 8.5 MHz) (fig. S5A): (i) Butterworth high-pass filter: lower cutoff frequency set to 1 MHz; (ii) Ram-Lak windowed ramp filter: linear within the bandwidth, with an abrupt cutoff at the frequency limit; (iii) Shepp-Logan windowed ramp filter: smooth enhancement within the bandwidth, with an abrupt cutoff; (iv) cosine windowed ramp filter: half-period cosine shape within the bandwidth, tapering to zero at both ends; and (v) hamming windowed ramp filter: smooth variation within the bandwidth (central frequency of 3.5 MHz), with a tapered attenuation near the cutoff frequency.

By comparing the raw signals with the outputs of the windowed filters, we found that the waveforms remained comparable, without yielding a noticeably “sharper” waveform (fig. S5, B to E). In the large FOV, the differences among the filters appeared minimal (fig. S5, F to J). However, in the small FOV, it was evident that the Butterworth high-pass filter provided the smoothest image, whereas the Ram-Lak window introduced fragmented patchy artifacts, which correspond to edge ringing (Gibbs phenomenon) (94) caused by abrupt spectral truncation (fig. S5, F to J). This phenomenon exists to varying degrees in all windowed ramp filters. However, it is weakest in the Hamming window, likely due to its specific filter shape design.

To perform a quantitative analysis, we approximately characterize the spatial resolution using the cross section of the hair and assessed the severity of ringing artifacts using a profile along the hair. Calculation of FWHM revealed that the feature width with high-pass filter was slightly inferior to that with the windowed ramp filters (fig. S5K). Regarding the SD, the high-pass filter yielded the lowest SD, indicating the smoothest image with minimal ringing, followed by the Hamming window, whereas the SDs of the other windowed filters were all higher than that of the Hamming window (fig. S5L). Therefore, we selected the Hamming-windowed ramp filter (central frequency of 3.5 MHz, upper cutoff frequency of 8.5 MHz) for data preprocessing as it achieves the optimal trade-off between spatial resolution and ringing artifact suppression.

Phantom and in vivo experiments

To provide a more concrete understanding of the spatial relationship between the transducer and the sample, as well as the overall system architecture, SolidWorks CAD model views of the system design are provided (fig. S15, A to D). For the spatial resolution assessment of 3D-ADPAT, a phantom was created by embedding a 10-μm-diameter tungsten wire in a cross configuration within a 2.5% (w/v) agar block. The leaf-vein phantom was prepared using black dye and subsequently embedding it in 2.5% (w/v) agar. The deep-tissue imaging phantom was fabricated by dispersing nickel-coated polystyrene conductive microspheres (100-μm diameter, Jiangsu Zhichuan Technology Co. Ltd.) in 2.5% (w/v) agar at a volume density of ~10 spheres/mm³. The hair phantoms, microsphere phantoms, and sucrose-agar phantoms used in the subsequent comparative experiments were all prepared using similar methods.

We included actual photographs of the system, including the optical illumination area and the in vivo animal immobilization setup, to verify the experimental configuration (fig. S15, E to G). Furthermore, supplementary movies showcasing the 3D-ADPAT system in operation during in vivo imaging sessions are provided to offer a more intuitive demonstration of the scanning and DAQ process (movies S7 and S8).

All animal experiments were performed in accordance with the principles outlined in the Guide for the Care and Use of Laboratory Animals of Hangzhou Medical College. All animal studies and experimental protocols were approved by the Institutional Animal Care and Use Committee at Hangzhou Medical College (approval no. SYXK2019-0011). An adult, 10-week-old female nude mouse (NU/NU, Charles River Co.) was used for the experiments. To minimize interference from GI signals, the mouse underwent a 12-hour fasting period before imaging to clear digestive residue. A small amount of depilatory cream was applied to the torso to remove fuzz from the skin. The animal’s torso was acoustically coupled to the sample stage using a thin layer of ultrasound gel. The convenient design of the system allowed for flexible adjustment of the animal’s posture (e.g., side-lying or prone). Throughout the imaging process, the animal was kept under anesthesia with 1.5% vaporized isoflurane delivered via a custom breathing tube fixed to the sample stage holder. The laser energy density incident on the animal’s skin (1064 nm, 22 mJ/cm²; 532 nm, 10 mJ/cm²) was well within the safety limits set by the American National Standards Institute. The animal recovered quickly after the imaging session and exhibited normal vital signs and activities in subsequent observations.

For in vivo probe tracking, 12-week-old mice (NU/NU, nude mouse, Charles River Co.) were used. A1094 (JHE0001, Suzhou Zhuoxinya Technology Co. Ltd., China) was selected as the small-molecule tracer. To prepare an A1094 solution, 0.3 mg of solid A1094 was first dissolved in 20 μl of dimethyl sulfoxide. Subsequently, 7 ml of distilled water was added, and the mixture was subjected to ultrasonication for 5 min to ensure complete dissolution. The injection dose of A1094 was 1 μmol per kilogram of body weight, and calculations indicated that ~0.4 ml of the prepared solution was required for injection. During the experiment, the animals were maintained under light anesthesia with 1.2% isoflurane. A 29-gauge indwelling tail vein catheter was connected to a syringe pump. Baseline imaging was performed first. Then, the pump was activated and the injection was completed within 90 s, with the animal remaining stationary throughout this operation. The mice behaved normally after the 8-hour probe tracking tests.

Acknowledgments

Funding:

This work was supported by the Key R&D Program of Zhejiang (no. 2024SSYS0014), Zhejiang Provincial Key Research and Development Program (no. 2021C03052), and Key Research Project of Zhejiang Lab (no. 2020MC0AD02), all awarded to J.S.; and by the National Natural Science Youth Foundation of China (no. 12304529) awarded to Y.R.

Author contributions:

Conceptualization: J.S. and X.W. Methodology: X.W., Y.X., X.G., Y.M., and R.W. Software: X.W., Y.W., and X.L. Resources: X.W., Y.X., and Y.W. Investigation: X.W., Y.X., Y.R., X.Y., F.M., Y.S., D.Z., X.G., and Z.X. Formal analysis: X.W. and Y.X. Visualization: X.W. Supervision: J.S. Writing—original draft: All authors. Writing—review and editing: X.W., Y.X., F.M., and J.S.

Competing interests:

The authors declare that they have no competing interests.

Data, code, and materials availability:

All data and code needed to evaluate and reproduce the results in the paper are present in the paper and/or the Supplementary Materials. The related photoacoustic data and reconstruction programs have been deposited in Figshare under collection DOI: 10.6084/m9.figshare.c.8397414. This study did not generate new materials.

Supplementary Materials

The PDF file includes:

Legends for movies S1 to S8

Figs. S1 to S15

Table S1

sciadv.aeb7362_sm.pdf^{(21.3MB, pdf)}

Other Supplementary Material for this manuscript includes the following:

Movies S1 to S8

sciadv.aeb7362_movies_s1_to_s8.zip^{(22.4MB, zip)}

REFERENCES

1.Xu M., Wang L. V., Photoacoustic imaging in biomedicine. Rev. Sci. Instrum. 77, 041101 (2006). [Google Scholar]
2.Wang L. V., Yao J., A practical guide to photoacoustic tomography in the life sciences. Nat. Methods 13, 627–638 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
3.Menozzi L., Yao J., Deep tissue photoacoustic imaging with light and sound. npj Imaging 2, 44 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
4.Hu P., Li L., Lin L., Wang L. V., Spatiotemporal antialiasing in photoacoustic computed tomography. IEEE Trans. Med. Imaging 39, 3535–3547 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
5.Luo X., Jiang J., Wu H., Li M., Wang B., The influences of finite aperture size in photoacoustic computed tomography. Ultrasonics 133, 107042 (2023). [DOI] [PubMed] [Google Scholar]
6.Qi L., Wu J., Li X., Zhang S., Huang S., Feng Q., Chen W., Photoacoustic tomography image restoration with measured spatially variant point spread functions. IEEE Trans. Med. Imaging 40, 2318–2328 (2021). [DOI] [PubMed] [Google Scholar]
7.Lin L., Hu P., Tong X., Na S., Cao R., Yuan X., Garrett D. C., Shi J., Maslov K., Wang L. V., High-speed three-dimensional photoacoustic computed tomography for preclinical research and clinical translation. Nat. Commun. 12, 882 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
8.Lv J., Lan H., Qin A., Sun T., Shao D., Gao F., Yao J., Avanaki K., Nie L., Dynamic synthetic-scanning photoacoustic tracking monitors hepatic and renal clearance pathway of exogeneous probes in vivo. Light Sci. Appl. 13, 304 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
9.Gao R., Chen T., Ren Y., Liu L., Chen N., Wong K. K. Y., Song L., Ma X., Liu C., Restoring the imaging quality of circular transducer array-based PACT using synthetic aperture focusing technique integrated with 2nd-derivative-based back projection scheme. Photoacoustics 32, 100537 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
10.Nagaoka R., Tabata T., Takagi R., Yoshizawa S., Umemura S. I., Saijo Y., Development of real-time 3-D photoacoustic imaging system employing spherically curved array transducer. IEEE Trans. Ultrason. Ferroelectr. Freq. Control 64, 1223–1233 (2017). [DOI] [PubMed] [Google Scholar]
11.Wissmeyer G., Pleitez M. A., Rosenthal A., Ntziachristos V., Looking at sound: Optoacoustics with all-optical ultrasound detection. Light Sci. Appl. 7, 53 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
12.Xu M., Wang L. V., Universal back-projection algorithm for photoacoustic computed tomography. Phys. Rev. E 71, 016706 (2005). [DOI] [PubMed] [Google Scholar]
13.Dean-Ben X. L., Buehler A., Ntziachristos V., Razansky D., Accurate model-based reconstruction algorithm for three-dimensional optoacoustic tomography. IEEE Trans. Med. Imaging 31, 1922–1928 (2012). [DOI] [PubMed] [Google Scholar]
14.Jin H., Zhang R., Liu S., Zheng Y., Fast and high-resolution three-dimensional hybrid-domain photoacoustic imaging incorporating analytical-focused transducer beam amplitude. IEEE Trans. Med. Imaging 38, 2926–2936 (2019). [DOI] [PubMed] [Google Scholar]
15.Li M.-L., Tseng Y.-C., Cheng C.-C., Model-based correction of finite aperture effect in photoacoustic tomography. Opt. Express 18, 26285–26292 (2010). [DOI] [PubMed] [Google Scholar]
16.Tang Y., Lesniak W. G., Gao S., Wu Y., Pomper M. G., Zhang H. K., Beyond synthetic aperture focusing: Deconvolution-based elevation resolution enhancement using simulated point spread function for linear array-based three-dimensional photoacoustic imaging. Biomed. Opt. Express 15, 1847–1860 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
17.Roitner H., Haltmeier M., Nuster R., O’Leary D. P., Berer T., Paltauf G., Grün H., Burgholzer P., Deblurring algorithms accounting for the finite detector size in photoacoustic tomography. J. Biomed. Opt. 19, 056011 (2014). [DOI] [PubMed] [Google Scholar]
18.Yang C., Lan H., Gao F., Gao F., Review of deep learning for photoacoustic imaging. Photoacoustics 21, 100215 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
19.Gröhl J., Schellenberg M., Dreher K., Maier-Hein L., Deep learning for biomedical photoacoustic imaging: A review. Photoacoustics 22, 100241 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
20.Chen B., Chen Y., Ma C., Photothermally tunable Fabry-Pérot fiber interferometer for photoacoustic mesoscopy. Biomed. Opt. Express 11, 2607–2618 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
21.Jathoul A. P., Laufer J., Ogunlade O., Treeby B., Cox B., Zhang E., Johnson P., Pizzey A. R., Philip B., Marafioti T., Lythgoe M. F., Pedley R. B., Pule M. A., Beard P., Deep in vivo photoacoustic imaging of mammalian tissues using a tyrosinase-based genetic reporter. Nat. Photonics 9, 239–246 (2015). [Google Scholar]
22.Huynh N., Zhang E., Francies O., Kuklis F., Allen T., Zhu J., Abeyakoon O., Lucka F., Betcke M., Jaros J., Arridge S., Cox B., Plumb A. A., Beard P., A fast all-optical 3D photoacoustic scanner for clinical vascular imaging. Nat. Biomed. Eng. 9, 638–655 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
23.Yang X., Li M.-L., Wang L. V., Ring-based ultrasonic virtual point detector with applications to photoacoustic tomography. Appl. Phys. Lett. 90, 251103 (2007). [Google Scholar]
24.Yang X., Wang L. V., Photoacoustic tomography of a rat cerebral cortex with a ring-based ultrasonic virtual point detector. J. Biomed. Opt. 12, 060507 (2007). [DOI] [PubMed] [Google Scholar]
25.Li M.-L., Zhang H. F., Maslov K., Stoica G., Wang L. V., Improved in vivo photoacoustic microscopy based on a virtual-detector concept. Opt. Lett. 31, 474–476 (2006). [DOI] [PubMed] [Google Scholar]
26.Li C., Wang L. V., Photoacoustic tomography of the mouse cerebral cortex with a high-numerical-aperture-based virtual point detector. J. Biomed. Opt. 14, 024047 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Nie L., Guo Z., Wang L. V., Photoacoustic tomography of monkey brain using virtual point ultrasonic transducers. J. Biomed. Opt. 16, 076005 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
28.Li C., Ku G., Wang L. V., Negative lens concept for photoacoustic tomography. Phys. Rev. E 78, 021901 (2008). [DOI] [PMC free article] [PubMed] [Google Scholar]
29.Hakakzadeh S., Mozaffarzadeh M., Mostafavi S. M., Kavehvash Z., Rajendran P., Verweij M., de Jong N., Pramanik M., Multi-angle data acquisition to compensate transducer finite size in photoacoustic tomography. Photoacoustics 27, 100373 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
30.Yoon C., Kang J., Song T.-k., Chang J. H., Elevational synthetic aperture focusing for three-dimensional photoacoustic imaging using a clinical one-dimensional array transducer. IEEE Trans. Biomed. Eng. 69, 2817–2825 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
31.Wang Y., Wang D., Zhang Y., Geng J., Lovell J. F., Xia J., Slit-enabled linear-array photoacoustic tomography with near isotropic spatial resolution in three dimensions. Opt. Lett. 41, 127–130 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
32.Wang Y., Wang D., Hubbell R., Xia J., Second generation slit-based photoacoustic tomography system for vascular imaging in human. J. Biophotonics 10, 799–804 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
33.Zheng W., Huang C., Zhang H., Xia J., Slit-based photoacoustic tomography with co-planar light illumination and acoustic detection for high-resolution vascular imaging in human using a linear transducer array. Biomed. Eng. Lett. 12, 125–133 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
34.Menozzi L., Vu T., Canning A. J., Rawtani H., Taboada C., Abi Antoun M. E., Ma C., Delia J., Nguyen V. T., Cho S. W., Chen J., Charity T., Xu Y., Tran P., Xia J., Palmer G. M., Vo-Dinh T., Feng L., Yao J., Three-dimensional diffractive acoustic tomography. Nat. Commun. 16, 1149 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
35.L. W. Schmerr, Fundamentals of Ultrasonic Nondestructive Evaluation (Springer, 2016), vol. 122. [Google Scholar]
36.Huang Y., Liu H., Chen X., Zhang Y., Zhao X., Huang W., Mou Z., Wang C., Zhang K., Li Z., Small-molecule absorber A1094 as a stable and fast-clearing NIR-II imaging agent. ChemMedChem 16, 2497–2503 (2021). [DOI] [PubMed] [Google Scholar]
37.Matsui O., Kadoya M., Kameyama T., Yoshikawa J., Takashima T., Nakanuma Y., Unoura M., Kobayashi K., Izumi R., Ida M., Benign and malignant nodules in cirrhotic livers: Distinction based on blood supply. Radiology 178, 493–497 (1991). [DOI] [PubMed] [Google Scholar]
38.Sänger C., Schenk A., Schwen L. O., Wang L., Gremse F., Zafarnia S., Kiessling F., Xie C., Wei W., Richter B., Dirsch O., Dahmen U., Intrahepatic vascular anatomy in rats and mice—variations and surgical implications. PLOS ONE 10, e0141798 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
39.Zhang X.-D., Wu H.-Y., Wu D., Wang Y.-Y., Chang J.-H., Zhai Z.-B., Meng A.-M., Liu P.-X., Zhang L.-A., Fan F.-Y., Toxicologic effects of gold nanoparticles in vivo by different administration routes. Int. J. Nanomedicine 5, 771–781 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
40.Poon W., Zhang Y.-N., Ouyang B., Kingston B. R., Wu J. L. Y., Wilhelm S., Chan W. C. W., Elimination pathways of nanoparticles. ACS nano 13, 5785–5798 (2019). [DOI] [PubMed] [Google Scholar]
41.Keinert B., Innmann M., Sänger M., Stamminger M., Spherical fibonacci mapping. ACM Trans. Graph. 34, 1–7 (2015). [Google Scholar]
42.Tong X., Lin L., Hu P., Cao R., Zhang Y., Olick-Gibson J., Wang L. V., Non-invasive 3D photoacoustic tomography of angiographic anatomy and hemodynamics of fatty livers in rats. Adv. Sci. 10, e2205759 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
43.Lin L., Tong X., Cavallero S., Zhang Y., Na S., Cao R., Hsiai T. K., Wang L. V., Non-invasive photoacoustic computed tomography of rat heart anatomy and function. Light Sci. Appl. 12, 12 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
44.Yao J., Wang L. V., Photoacoustic microscopy. Laser Photonics Rev. 7, 758–778 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
45.Jeon S., Kim J., Lee D., Baik J. W., Kim C., Review on practical photoacoustic microscopy. Photoacoustics 15, 100141 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
46.Cao R., Zhao J., Li L., du L., Zhang Y., Luo Y., Jiang L., Davis S., Zhou Q., de la Zerda A., Wang L. V., Optical-resolution photoacoustic microscopy with a needle-shaped beam. Nat. Photonics 17, 89–95 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
47.Choi S., Yang J., Lee S. Y., Kim J., Lee J., Kim W. J., Lee S., Kim C., Deep learning enhances multiparametric dynamic volumetric photoacoustic computed tomography in vivo (DL-PACT). Adv. Sci. 10, e2202089 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
48.Asao Y., Nagae K., Miyasaka K., Sekiguchi H., Aiso S., Watanabe S., Sato M., Kizaka-Kondoh S., Nakajima Y., Kishi K., Yagi T., In vivo label-free observation of tumor-related blood vessels in small animals using a newly designed photoacoustic 3D imaging system. Ultrason. Imaging 44, 96–104 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
49.Yang J.-M., Favazza C., Chen R., Yao J., Cai X., Maslov K., Zhou Q., Shung K. K., Wang L. V., Simultaneous functional photoacoustic and ultrasonic endoscopy of internal organs in vivo. Nat. Med. 18, 1297–1302 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
50.Guo H., Li Y., Qi W., Xi L., Photoacoustic endoscopy: A progress review. J. Biophotonics 13, e202000217 (2020). [DOI] [PubMed] [Google Scholar]
51.Xia J., Guo Z., Maslov K. I., Wang L. V., Aguirre A., Zhu Q., Percival C., Three-dimensional photoacoustic tomography based on the focal-line concept. J. Biomed. Opt. 16, 090505 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
52.Wang D., Wang Y., Zhou Y., Lovell J. F., Xia J., Coherent-weighted three-dimensional image reconstruction in linear-array-based photoacoustic tomography. Biomed. Opt. Express 7, 1957–1965 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
53.Xie D., Dong W., Zheng J., Tian C., Spatially-variant image deconvolution for photoacoustic tomography. Opt. Express 31, 21641–21657 (2023). [DOI] [PubMed] [Google Scholar]
54.Chowdhury K. B., Prakash J., Karlas A., Jüstel D., Ntziachristos V., A synthetic total impulse response characterization method for correction of hand-held optoacoustic images. IEEE Trans. Med. Imaging 39, 3218–3230 (2020). [DOI] [PubMed] [Google Scholar]
55.Richardson W. H., Bayesian-based iterative method of image restoration. J. Opt. Soc. Am. 62, 55–59 (1972). [Google Scholar]
56.Lucy L. B., An iterative technique for the rectification of observed distributions. Astron. J. 79, 745 (1974). [Google Scholar]
57.M. Bertero, P. Boccacci, C. De Mol, Introduction to Inverse Problems in Imaging (CRC Press, 2021). [Google Scholar]
58.Rosenthal A., Ntziachristos V., Razansky D., Acoustic inversion in optoacoustic tomography: A review. Curr. Med. Imaging 9, 318–336 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
59.Savakus H., Klicker K., Newnham R., PZT-epoxy piezoelectric transducers: A simplified fabrication procedure. Mater. Res. Bull. 16, 677–680 (1981). [Google Scholar]
60.Huang K., Fu P., Zhu H., Feng J., Zhang L., Wang B., Lu Y., Zhang D., Yao M., Chen L., Ying Y., Chen J., Li X., Wu Y., Xiong W., Li J., Wu Y., Sun J., Zhang H., Lin L., High-speed photoacoustic and ultrasonic computed tomography of the breast tumor for early diagnosis with enhanced accuracy. Sci. Adv. 11, eadz2046 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
61.Kalva S. K., Özsoy C., Nozdriukhin D., Tiemann S., Tang L., Deán-Ben X. L., Razansky D., Toward noninvasive optoacoustic imaging of whole-heart dynamics in mice. Light Sci. Appl. 14, 391 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
62.Ali R., Jin G., Singh M., Mitcham T., Duric N., 3D frequency-domain full waveform inversion for whole-breast imaging with a multi-row ring array. IEEE Open J. Ultrason. Ferroelectr. Freq. Control 5, 77–81 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
63.Cox B., Laufer J. G., Arridge S. R., Beard P. C., Quantitative spectroscopic photoacoustic imaging: A review. J. Biomed. Opt. 17, 061202 (2012). [DOI] [PubMed] [Google Scholar]
64.Zeng L., Ma G., Lin J., Huang P., Photoacoustic probes for molecular detection: Recent advances and perspectives. Small 14, 1800782 (2018). [DOI] [PubMed] [Google Scholar]
65.Kim M., Han J. H., Ahn J., Kim E., Bang C. H., Kim C., Lee J. H., Choi W., In vivo 3D photoacoustic and ultrasound analysis of hypopigmented skin lesions: A pilot study. Photoacoustics 43, 100705 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
66.Huang C., Cheng Y., Zheng W., Bing R. W., Zhang H., Komornicki I., Harris L. M., Arany P. R., Chakraborty S., Zhou Q., Xu W., Xia J., Dual-scan photoacoustic tomography for the imaging of vascular structure on foot. IEEE Trans. Ultrason. Ferroelectr. Freq. Control 70, 1703–1713 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
67.Jiang H., Zhao A., Yang Q., Yan X., Wang T., Wang Y., Jia N., Wang J., Wu G., Yue Y., Luo S., Wang H., Ren L., Chen S., Liu P., Yao G., Yang W., Song S., Li X., He K., Huang G., Towards expert-level autonomous carotid ultrasonography with large-scale learning-based robotic system. Nat. Commun. 16, 7893 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
68.Ahn J., Lee J., Kim K., Bae J. S., Jung C. K., Kim M., Choi W., Park B., Lim D. J., Kim C., Smarter biopsy decisions in thyroid nodules via dual-modal photoacoustic and ultrasound imaging. Sci. Adv. 11, eady6173 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
69.Estrada H., Chen Y., Lemaire T., Davoudi N., Özbek A., Parduzi Q., Shoham S., Razansky D., Holographic transcranial ultrasound neuromodulation enhances stimulation efficacy by cooperatively recruiting distributed brain circuits. Nat. Biomed. Eng. 10, 6–15 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]
70.Cudeiro-Blanco J., Cueto C., Bates O., Strong G., Robins T., Toulemonde M., Warner M., Tang M. X., Agudo O. C., Guasch L., Design and construction of a low-frequency ultrasound acquisition device for 2-D brain imaging using full-waveform inversion. Ultrasound Med. Biol. 48, 1995–2008 (2022). [DOI] [PubMed] [Google Scholar]
71.Li Y., Shi Q., Li Y., Song X., Liu C., Ta D., Wang W., High-resolution bone microstructure imaging based on ultrasonic frequency-domain full-waveform inversion. Chinese Phys. B 30, 014302 (2021). [Google Scholar]
72.Chen J., Ye X., Gao S., Chen Y., Zhao Y., Huang C., Qiu K., Zhu S., Li T., Planar wide-angle-imaging camera enabled by metalens array. Optica 9, 431–437 (2022). [Google Scholar]
73.J. Chen, Y. Zhang, Z. Zhang, “Deep learning enhanced color super-resolution imaging in metalens-integrated camer,” in 2024 4th International Conference on Computer Communication and Artificial Intelligence (CCAI) (IEEE, 2024), pp. 94–99. [Google Scholar]
74.Zhang Y., Wu Y., Huang C., Zhou Z. W., Li M., Zhang Z., Chen J., Deep-learning enhanced high-quality imaging in metalens-integrated camera. Opt. Lett. 49, 2853–2856 (2024). [DOI] [PubMed] [Google Scholar]
75.Davoudi N., Deán-Ben X. L., Razansky D., Deep learning optoacoustic tomography with sparse data. Nat. Mach. Intell. 1, 453–460 (2019). [Google Scholar]
76.Yuan Z., Yang D., Wang W., Zhao J., Liang Y., Self super-resolution of optical coherence tomography images based on deep learning. Opt. Express 31, 27566–27581 (2023). [DOI] [PubMed] [Google Scholar]
77.Deng K., Luo Y., Zuo H., Chen Y., Gu L., Liu M., Lan H., Luo J., Ma C., Self-supervised upsampling for reconstructions with generalized enhancement in photoacoustic computed tomography. IEEE Trans. Med. Imaging 44, 5117–5127 (2025). [DOI] [PubMed] [Google Scholar]
78.Y. Mansour, R. Heckel, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (IEEE, 2023), pp. 14018–14027. [Google Scholar]
79.Jensen J. A., Field: A program for simulating ultrasound systems. Med. Biol. Eng. Comput. 34, 351–353 (1997). [Google Scholar]
80.Jensen J. A., Svendsen N. B., Calculation of pressure fields from arbitrarily shaped, apodized, and excited ultrasound transducers. IEEE Trans. Ultrason. Ferroelectr. Freq. Control 39, 262–267 (1992). [DOI] [PubMed] [Google Scholar]
81.Harris G. R., Review of transient field theory for a baffled planar piston. J. Acoust. Soc. Am. 70, 10–20 (1981). [Google Scholar]
82.F. Duck, Physical Properties of Tissues: A Comprehensive Reference Book (Academic Press, 2013). [Google Scholar]
83.Hynynen K., Clement G., Clinical applications of focused ultrasound—The brain. Int. J. Hyperthermia 23, 193–202 (2007). [DOI] [PubMed] [Google Scholar]
84.Stepanishen P. R., Transient radiation from pistons in an infinite planar baffle. J. Acoust. Soc. Am. 49, 1629–1638 (1971). [Google Scholar]
85.Culjat M. O., Goldenberg D., Tewari P., Singh R. S., A review of tissue substitutes for ultrasound imaging. Ultrasound Med. Biol. 36, 861–873 (2010). [DOI] [PubMed] [Google Scholar]
86.Marczak W., Water as a standard in the measurements of speed of sound in liquids. J. Acoust. Soc. Am. 102, 2776–2779 (1997). [Google Scholar]
87.Shi G., Lu H., Hui H., Tian J., Benefit from public unlabeled data: A Frangi filter-based pretraining network for 3D cerebrovascular segmentation. Med. Image Anal. 101, 103442 (2025). [DOI] [PubMed] [Google Scholar]
88.Campbell I. C., Coudrillier B., Mensah J., Abel R. L., Ethier C. R., Automated segmentation of the lamina cribrosa using Frangi’s filter: A novel approach for rapid identification of tissue volume fraction and beam orientation in a trabeculated structure in the eye. J. R. Soc. Interface 12, 20141009 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
89.Merčep E., Herraiz J. L., Deán-Ben X. L., Razansky D., Transmission–reflection optoacoustic ultrasound (TROPUS) computed tomography of small animals. Light Sci. Appl. 8, 18 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
90.Li L., Zhu L., Ma C., Lin L., Yao J., Wang L., Maslov K., Zhang R., Chen W., Shi J., Wang L. V., Single-impulse panoramic photoacoustic computed tomography of small-animal whole-body dynamics at high spatiotemporal resolution. Nat. Biomed. Eng. 1, 0071 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
91.Ramachandran G., Lakshminarayanan A., Three-dimensional reconstruction from radiographs and electron micrographs: Application of convolutions instead of Fourier transforms. Proc. Natl. Acad. Sci. U.S.A. 68, 2236–2240 (1971). [DOI] [PMC free article] [PubMed] [Google Scholar]
92.Shepp L. A., Logan B. F., The Fourier reconstruction of a head section. IEEE Trans. Nucl. Sci. 21, 21–43 (1974). [Google Scholar]
93.A. C. Kak, M. Slaney, Principles of Computerized Tomographic Imaging (SIAM, 2001). [Google Scholar]
94.Gottlieb D., Shu C.-W., On the Gibbs phenomenon and its resolution. SIAM Rev. 39, 644–668 (1997). [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Legends for movies S1 to S8

Figs. S1 to S15

Table S1

sciadv.aeb7362_sm.pdf^{(21.3MB, pdf)}

Movies S1 to S8

sciadv.aeb7362_movies_s1_to_s8.zip^{(22.4MB, zip)}

Data Availability Statement

[R1] 1.Xu M., Wang L. V., Photoacoustic imaging in biomedicine. Rev. Sci. Instrum. 77, 041101 (2006). [Google Scholar]

[R2] 2.Wang L. V., Yao J., A practical guide to photoacoustic tomography in the life sciences. Nat. Methods 13, 627–638 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R3] 3.Menozzi L., Yao J., Deep tissue photoacoustic imaging with light and sound. npj Imaging 2, 44 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R4] 4.Hu P., Li L., Lin L., Wang L. V., Spatiotemporal antialiasing in photoacoustic computed tomography. IEEE Trans. Med. Imaging 39, 3535–3547 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R5] 5.Luo X., Jiang J., Wu H., Li M., Wang B., The influences of finite aperture size in photoacoustic computed tomography. Ultrasonics 133, 107042 (2023). [DOI] [PubMed] [Google Scholar]

[R6] 6.Qi L., Wu J., Li X., Zhang S., Huang S., Feng Q., Chen W., Photoacoustic tomography image restoration with measured spatially variant point spread functions. IEEE Trans. Med. Imaging 40, 2318–2328 (2021). [DOI] [PubMed] [Google Scholar]

[R7] 7.Lin L., Hu P., Tong X., Na S., Cao R., Yuan X., Garrett D. C., Shi J., Maslov K., Wang L. V., High-speed three-dimensional photoacoustic computed tomography for preclinical research and clinical translation. Nat. Commun. 12, 882 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R8] 8.Lv J., Lan H., Qin A., Sun T., Shao D., Gao F., Yao J., Avanaki K., Nie L., Dynamic synthetic-scanning photoacoustic tracking monitors hepatic and renal clearance pathway of exogeneous probes in vivo. Light Sci. Appl. 13, 304 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R9] 9.Gao R., Chen T., Ren Y., Liu L., Chen N., Wong K. K. Y., Song L., Ma X., Liu C., Restoring the imaging quality of circular transducer array-based PACT using synthetic aperture focusing technique integrated with 2nd-derivative-based back projection scheme. Photoacoustics 32, 100537 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R10] 10.Nagaoka R., Tabata T., Takagi R., Yoshizawa S., Umemura S. I., Saijo Y., Development of real-time 3-D photoacoustic imaging system employing spherically curved array transducer. IEEE Trans. Ultrason. Ferroelectr. Freq. Control 64, 1223–1233 (2017). [DOI] [PubMed] [Google Scholar]

[R11] 11.Wissmeyer G., Pleitez M. A., Rosenthal A., Ntziachristos V., Looking at sound: Optoacoustics with all-optical ultrasound detection. Light Sci. Appl. 7, 53 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R12] 12.Xu M., Wang L. V., Universal back-projection algorithm for photoacoustic computed tomography. Phys. Rev. E 71, 016706 (2005). [DOI] [PubMed] [Google Scholar]

[R13] 13.Dean-Ben X. L., Buehler A., Ntziachristos V., Razansky D., Accurate model-based reconstruction algorithm for three-dimensional optoacoustic tomography. IEEE Trans. Med. Imaging 31, 1922–1928 (2012). [DOI] [PubMed] [Google Scholar]

[R14] 14.Jin H., Zhang R., Liu S., Zheng Y., Fast and high-resolution three-dimensional hybrid-domain photoacoustic imaging incorporating analytical-focused transducer beam amplitude. IEEE Trans. Med. Imaging 38, 2926–2936 (2019). [DOI] [PubMed] [Google Scholar]

[R15] 15.Li M.-L., Tseng Y.-C., Cheng C.-C., Model-based correction of finite aperture effect in photoacoustic tomography. Opt. Express 18, 26285–26292 (2010). [DOI] [PubMed] [Google Scholar]

[R16] 16.Tang Y., Lesniak W. G., Gao S., Wu Y., Pomper M. G., Zhang H. K., Beyond synthetic aperture focusing: Deconvolution-based elevation resolution enhancement using simulated point spread function for linear array-based three-dimensional photoacoustic imaging. Biomed. Opt. Express 15, 1847–1860 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R17] 17.Roitner H., Haltmeier M., Nuster R., O’Leary D. P., Berer T., Paltauf G., Grün H., Burgholzer P., Deblurring algorithms accounting for the finite detector size in photoacoustic tomography. J. Biomed. Opt. 19, 056011 (2014). [DOI] [PubMed] [Google Scholar]

[R18] 18.Yang C., Lan H., Gao F., Gao F., Review of deep learning for photoacoustic imaging. Photoacoustics 21, 100215 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R19] 19.Gröhl J., Schellenberg M., Dreher K., Maier-Hein L., Deep learning for biomedical photoacoustic imaging: A review. Photoacoustics 22, 100241 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R20] 20.Chen B., Chen Y., Ma C., Photothermally tunable Fabry-Pérot fiber interferometer for photoacoustic mesoscopy. Biomed. Opt. Express 11, 2607–2618 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R21] 21.Jathoul A. P., Laufer J., Ogunlade O., Treeby B., Cox B., Zhang E., Johnson P., Pizzey A. R., Philip B., Marafioti T., Lythgoe M. F., Pedley R. B., Pule M. A., Beard P., Deep in vivo photoacoustic imaging of mammalian tissues using a tyrosinase-based genetic reporter. Nat. Photonics 9, 239–246 (2015). [Google Scholar]

[R22] 22.Huynh N., Zhang E., Francies O., Kuklis F., Allen T., Zhu J., Abeyakoon O., Lucka F., Betcke M., Jaros J., Arridge S., Cox B., Plumb A. A., Beard P., A fast all-optical 3D photoacoustic scanner for clinical vascular imaging. Nat. Biomed. Eng. 9, 638–655 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R23] 23.Yang X., Li M.-L., Wang L. V., Ring-based ultrasonic virtual point detector with applications to photoacoustic tomography. Appl. Phys. Lett. 90, 251103 (2007). [Google Scholar]

[R24] 24.Yang X., Wang L. V., Photoacoustic tomography of a rat cerebral cortex with a ring-based ultrasonic virtual point detector. J. Biomed. Opt. 12, 060507 (2007). [DOI] [PubMed] [Google Scholar]

[R25] 25.Li M.-L., Zhang H. F., Maslov K., Stoica G., Wang L. V., Improved in vivo photoacoustic microscopy based on a virtual-detector concept. Opt. Lett. 31, 474–476 (2006). [DOI] [PubMed] [Google Scholar]

[R26] 26.Li C., Wang L. V., Photoacoustic tomography of the mouse cerebral cortex with a high-numerical-aperture-based virtual point detector. J. Biomed. Opt. 14, 024047 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R27] 27.Nie L., Guo Z., Wang L. V., Photoacoustic tomography of monkey brain using virtual point ultrasonic transducers. J. Biomed. Opt. 16, 076005 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R28] 28.Li C., Ku G., Wang L. V., Negative lens concept for photoacoustic tomography. Phys. Rev. E 78, 021901 (2008). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R29] 29.Hakakzadeh S., Mozaffarzadeh M., Mostafavi S. M., Kavehvash Z., Rajendran P., Verweij M., de Jong N., Pramanik M., Multi-angle data acquisition to compensate transducer finite size in photoacoustic tomography. Photoacoustics 27, 100373 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R30] 30.Yoon C., Kang J., Song T.-k., Chang J. H., Elevational synthetic aperture focusing for three-dimensional photoacoustic imaging using a clinical one-dimensional array transducer. IEEE Trans. Biomed. Eng. 69, 2817–2825 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R31] 31.Wang Y., Wang D., Zhang Y., Geng J., Lovell J. F., Xia J., Slit-enabled linear-array photoacoustic tomography with near isotropic spatial resolution in three dimensions. Opt. Lett. 41, 127–130 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R32] 32.Wang Y., Wang D., Hubbell R., Xia J., Second generation slit-based photoacoustic tomography system for vascular imaging in human. J. Biophotonics 10, 799–804 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R33] 33.Zheng W., Huang C., Zhang H., Xia J., Slit-based photoacoustic tomography with co-planar light illumination and acoustic detection for high-resolution vascular imaging in human using a linear transducer array. Biomed. Eng. Lett. 12, 125–133 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R34] 34.Menozzi L., Vu T., Canning A. J., Rawtani H., Taboada C., Abi Antoun M. E., Ma C., Delia J., Nguyen V. T., Cho S. W., Chen J., Charity T., Xu Y., Tran P., Xia J., Palmer G. M., Vo-Dinh T., Feng L., Yao J., Three-dimensional diffractive acoustic tomography. Nat. Commun. 16, 1149 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R35] 35.L. W. Schmerr, Fundamentals of Ultrasonic Nondestructive Evaluation (Springer, 2016), vol. 122. [Google Scholar]

[R36] 36.Huang Y., Liu H., Chen X., Zhang Y., Zhao X., Huang W., Mou Z., Wang C., Zhang K., Li Z., Small-molecule absorber A1094 as a stable and fast-clearing NIR-II imaging agent. ChemMedChem 16, 2497–2503 (2021). [DOI] [PubMed] [Google Scholar]

[R37] 37.Matsui O., Kadoya M., Kameyama T., Yoshikawa J., Takashima T., Nakanuma Y., Unoura M., Kobayashi K., Izumi R., Ida M., Benign and malignant nodules in cirrhotic livers: Distinction based on blood supply. Radiology 178, 493–497 (1991). [DOI] [PubMed] [Google Scholar]

[R38] 38.Sänger C., Schenk A., Schwen L. O., Wang L., Gremse F., Zafarnia S., Kiessling F., Xie C., Wei W., Richter B., Dirsch O., Dahmen U., Intrahepatic vascular anatomy in rats and mice—variations and surgical implications. PLOS ONE 10, e0141798 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R39] 39.Zhang X.-D., Wu H.-Y., Wu D., Wang Y.-Y., Chang J.-H., Zhai Z.-B., Meng A.-M., Liu P.-X., Zhang L.-A., Fan F.-Y., Toxicologic effects of gold nanoparticles in vivo by different administration routes. Int. J. Nanomedicine 5, 771–781 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R40] 40.Poon W., Zhang Y.-N., Ouyang B., Kingston B. R., Wu J. L. Y., Wilhelm S., Chan W. C. W., Elimination pathways of nanoparticles. ACS nano 13, 5785–5798 (2019). [DOI] [PubMed] [Google Scholar]

[R41] 41.Keinert B., Innmann M., Sänger M., Stamminger M., Spherical fibonacci mapping. ACM Trans. Graph. 34, 1–7 (2015). [Google Scholar]

[R42] 42.Tong X., Lin L., Hu P., Cao R., Zhang Y., Olick-Gibson J., Wang L. V., Non-invasive 3D photoacoustic tomography of angiographic anatomy and hemodynamics of fatty livers in rats. Adv. Sci. 10, e2205759 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R43] 43.Lin L., Tong X., Cavallero S., Zhang Y., Na S., Cao R., Hsiai T. K., Wang L. V., Non-invasive photoacoustic computed tomography of rat heart anatomy and function. Light Sci. Appl. 12, 12 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R44] 44.Yao J., Wang L. V., Photoacoustic microscopy. Laser Photonics Rev. 7, 758–778 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R45] 45.Jeon S., Kim J., Lee D., Baik J. W., Kim C., Review on practical photoacoustic microscopy. Photoacoustics 15, 100141 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R46] 46.Cao R., Zhao J., Li L., du L., Zhang Y., Luo Y., Jiang L., Davis S., Zhou Q., de la Zerda A., Wang L. V., Optical-resolution photoacoustic microscopy with a needle-shaped beam. Nat. Photonics 17, 89–95 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R47] 47.Choi S., Yang J., Lee S. Y., Kim J., Lee J., Kim W. J., Lee S., Kim C., Deep learning enhances multiparametric dynamic volumetric photoacoustic computed tomography in vivo (DL-PACT). Adv. Sci. 10, e2202089 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R48] 48.Asao Y., Nagae K., Miyasaka K., Sekiguchi H., Aiso S., Watanabe S., Sato M., Kizaka-Kondoh S., Nakajima Y., Kishi K., Yagi T., In vivo label-free observation of tumor-related blood vessels in small animals using a newly designed photoacoustic 3D imaging system. Ultrason. Imaging 44, 96–104 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R49] 49.Yang J.-M., Favazza C., Chen R., Yao J., Cai X., Maslov K., Zhou Q., Shung K. K., Wang L. V., Simultaneous functional photoacoustic and ultrasonic endoscopy of internal organs in vivo. Nat. Med. 18, 1297–1302 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R50] 50.Guo H., Li Y., Qi W., Xi L., Photoacoustic endoscopy: A progress review. J. Biophotonics 13, e202000217 (2020). [DOI] [PubMed] [Google Scholar]

[R51] 51.Xia J., Guo Z., Maslov K. I., Wang L. V., Aguirre A., Zhu Q., Percival C., Three-dimensional photoacoustic tomography based on the focal-line concept. J. Biomed. Opt. 16, 090505 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R52] 52.Wang D., Wang Y., Zhou Y., Lovell J. F., Xia J., Coherent-weighted three-dimensional image reconstruction in linear-array-based photoacoustic tomography. Biomed. Opt. Express 7, 1957–1965 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R53] 53.Xie D., Dong W., Zheng J., Tian C., Spatially-variant image deconvolution for photoacoustic tomography. Opt. Express 31, 21641–21657 (2023). [DOI] [PubMed] [Google Scholar]

[R54] 54.Chowdhury K. B., Prakash J., Karlas A., Jüstel D., Ntziachristos V., A synthetic total impulse response characterization method for correction of hand-held optoacoustic images. IEEE Trans. Med. Imaging 39, 3218–3230 (2020). [DOI] [PubMed] [Google Scholar]

[R55] 55.Richardson W. H., Bayesian-based iterative method of image restoration. J. Opt. Soc. Am. 62, 55–59 (1972). [Google Scholar]

[R56] 56.Lucy L. B., An iterative technique for the rectification of observed distributions. Astron. J. 79, 745 (1974). [Google Scholar]

[R57] 57.M. Bertero, P. Boccacci, C. De Mol, Introduction to Inverse Problems in Imaging (CRC Press, 2021). [Google Scholar]

[R58] 58.Rosenthal A., Ntziachristos V., Razansky D., Acoustic inversion in optoacoustic tomography: A review. Curr. Med. Imaging 9, 318–336 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R59] 59.Savakus H., Klicker K., Newnham R., PZT-epoxy piezoelectric transducers: A simplified fabrication procedure. Mater. Res. Bull. 16, 677–680 (1981). [Google Scholar]

[R60] 60.Huang K., Fu P., Zhu H., Feng J., Zhang L., Wang B., Lu Y., Zhang D., Yao M., Chen L., Ying Y., Chen J., Li X., Wu Y., Xiong W., Li J., Wu Y., Sun J., Zhang H., Lin L., High-speed photoacoustic and ultrasonic computed tomography of the breast tumor for early diagnosis with enhanced accuracy. Sci. Adv. 11, eadz2046 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R61] 61.Kalva S. K., Özsoy C., Nozdriukhin D., Tiemann S., Tang L., Deán-Ben X. L., Razansky D., Toward noninvasive optoacoustic imaging of whole-heart dynamics in mice. Light Sci. Appl. 14, 391 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R62] 62.Ali R., Jin G., Singh M., Mitcham T., Duric N., 3D frequency-domain full waveform inversion for whole-breast imaging with a multi-row ring array. IEEE Open J. Ultrason. Ferroelectr. Freq. Control 5, 77–81 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R63] 63.Cox B., Laufer J. G., Arridge S. R., Beard P. C., Quantitative spectroscopic photoacoustic imaging: A review. J. Biomed. Opt. 17, 061202 (2012). [DOI] [PubMed] [Google Scholar]

[R64] 64.Zeng L., Ma G., Lin J., Huang P., Photoacoustic probes for molecular detection: Recent advances and perspectives. Small 14, 1800782 (2018). [DOI] [PubMed] [Google Scholar]

[R65] 65.Kim M., Han J. H., Ahn J., Kim E., Bang C. H., Kim C., Lee J. H., Choi W., In vivo 3D photoacoustic and ultrasound analysis of hypopigmented skin lesions: A pilot study. Photoacoustics 43, 100705 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R66] 66.Huang C., Cheng Y., Zheng W., Bing R. W., Zhang H., Komornicki I., Harris L. M., Arany P. R., Chakraborty S., Zhou Q., Xu W., Xia J., Dual-scan photoacoustic tomography for the imaging of vascular structure on foot. IEEE Trans. Ultrason. Ferroelectr. Freq. Control 70, 1703–1713 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R67] 67.Jiang H., Zhao A., Yang Q., Yan X., Wang T., Wang Y., Jia N., Wang J., Wu G., Yue Y., Luo S., Wang H., Ren L., Chen S., Liu P., Yao G., Yang W., Song S., Li X., He K., Huang G., Towards expert-level autonomous carotid ultrasonography with large-scale learning-based robotic system. Nat. Commun. 16, 7893 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R68] 68.Ahn J., Lee J., Kim K., Bae J. S., Jung C. K., Kim M., Choi W., Park B., Lim D. J., Kim C., Smarter biopsy decisions in thyroid nodules via dual-modal photoacoustic and ultrasound imaging. Sci. Adv. 11, eady6173 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R69] 69.Estrada H., Chen Y., Lemaire T., Davoudi N., Özbek A., Parduzi Q., Shoham S., Razansky D., Holographic transcranial ultrasound neuromodulation enhances stimulation efficacy by cooperatively recruiting distributed brain circuits. Nat. Biomed. Eng. 10, 6–15 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R70] 70.Cudeiro-Blanco J., Cueto C., Bates O., Strong G., Robins T., Toulemonde M., Warner M., Tang M. X., Agudo O. C., Guasch L., Design and construction of a low-frequency ultrasound acquisition device for 2-D brain imaging using full-waveform inversion. Ultrasound Med. Biol. 48, 1995–2008 (2022). [DOI] [PubMed] [Google Scholar]

[R71] 71.Li Y., Shi Q., Li Y., Song X., Liu C., Ta D., Wang W., High-resolution bone microstructure imaging based on ultrasonic frequency-domain full-waveform inversion. Chinese Phys. B 30, 014302 (2021). [Google Scholar]

[R72] 72.Chen J., Ye X., Gao S., Chen Y., Zhao Y., Huang C., Qiu K., Zhu S., Li T., Planar wide-angle-imaging camera enabled by metalens array. Optica 9, 431–437 (2022). [Google Scholar]

[R73] 73.J. Chen, Y. Zhang, Z. Zhang, “Deep learning enhanced color super-resolution imaging in metalens-integrated camer,” in 2024 4th International Conference on Computer Communication and Artificial Intelligence (CCAI) (IEEE, 2024), pp. 94–99. [Google Scholar]

[R74] 74.Zhang Y., Wu Y., Huang C., Zhou Z. W., Li M., Zhang Z., Chen J., Deep-learning enhanced high-quality imaging in metalens-integrated camera. Opt. Lett. 49, 2853–2856 (2024). [DOI] [PubMed] [Google Scholar]

[R75] 75.Davoudi N., Deán-Ben X. L., Razansky D., Deep learning optoacoustic tomography with sparse data. Nat. Mach. Intell. 1, 453–460 (2019). [Google Scholar]

[R76] 76.Yuan Z., Yang D., Wang W., Zhao J., Liang Y., Self super-resolution of optical coherence tomography images based on deep learning. Opt. Express 31, 27566–27581 (2023). [DOI] [PubMed] [Google Scholar]

[R77] 77.Deng K., Luo Y., Zuo H., Chen Y., Gu L., Liu M., Lan H., Luo J., Ma C., Self-supervised upsampling for reconstructions with generalized enhancement in photoacoustic computed tomography. IEEE Trans. Med. Imaging 44, 5117–5127 (2025). [DOI] [PubMed] [Google Scholar]

[R78] 78.Y. Mansour, R. Heckel, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (IEEE, 2023), pp. 14018–14027. [Google Scholar]

[R79] 79.Jensen J. A., Field: A program for simulating ultrasound systems. Med. Biol. Eng. Comput. 34, 351–353 (1997). [Google Scholar]

[R80] 80.Jensen J. A., Svendsen N. B., Calculation of pressure fields from arbitrarily shaped, apodized, and excited ultrasound transducers. IEEE Trans. Ultrason. Ferroelectr. Freq. Control 39, 262–267 (1992). [DOI] [PubMed] [Google Scholar]

[R81] 81.Harris G. R., Review of transient field theory for a baffled planar piston. J. Acoust. Soc. Am. 70, 10–20 (1981). [Google Scholar]

[R82] 82.F. Duck, Physical Properties of Tissues: A Comprehensive Reference Book (Academic Press, 2013). [Google Scholar]

[R83] 83.Hynynen K., Clement G., Clinical applications of focused ultrasound—The brain. Int. J. Hyperthermia 23, 193–202 (2007). [DOI] [PubMed] [Google Scholar]

[R84] 84.Stepanishen P. R., Transient radiation from pistons in an infinite planar baffle. J. Acoust. Soc. Am. 49, 1629–1638 (1971). [Google Scholar]

[R85] 85.Culjat M. O., Goldenberg D., Tewari P., Singh R. S., A review of tissue substitutes for ultrasound imaging. Ultrasound Med. Biol. 36, 861–873 (2010). [DOI] [PubMed] [Google Scholar]

[R86] 86.Marczak W., Water as a standard in the measurements of speed of sound in liquids. J. Acoust. Soc. Am. 102, 2776–2779 (1997). [Google Scholar]

[R87] 87.Shi G., Lu H., Hui H., Tian J., Benefit from public unlabeled data: A Frangi filter-based pretraining network for 3D cerebrovascular segmentation. Med. Image Anal. 101, 103442 (2025). [DOI] [PubMed] [Google Scholar]

[R88] 88.Campbell I. C., Coudrillier B., Mensah J., Abel R. L., Ethier C. R., Automated segmentation of the lamina cribrosa using Frangi’s filter: A novel approach for rapid identification of tissue volume fraction and beam orientation in a trabeculated structure in the eye. J. R. Soc. Interface 12, 20141009 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R89] 89.Merčep E., Herraiz J. L., Deán-Ben X. L., Razansky D., Transmission–reflection optoacoustic ultrasound (TROPUS) computed tomography of small animals. Light Sci. Appl. 8, 18 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R90] 90.Li L., Zhu L., Ma C., Lin L., Yao J., Wang L., Maslov K., Zhang R., Chen W., Shi J., Wang L. V., Single-impulse panoramic photoacoustic computed tomography of small-animal whole-body dynamics at high spatiotemporal resolution. Nat. Biomed. Eng. 1, 0071 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R91] 91.Ramachandran G., Lakshminarayanan A., Three-dimensional reconstruction from radiographs and electron micrographs: Application of convolutions instead of Fourier transforms. Proc. Natl. Acad. Sci. U.S.A. 68, 2236–2240 (1971). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R92] 92.Shepp L. A., Logan B. F., The Fourier reconstruction of a head section. IEEE Trans. Nucl. Sci. 21, 21–43 (1974). [Google Scholar]

[R93] 93.A. C. Kak, M. Slaney, Principles of Computerized Tomographic Imaging (SIAM, 2001). [Google Scholar]

[R94] 94.Gottlieb D., Shu C.-W., On the Gibbs phenomenon and its resolution. SIAM Rev. 39, 644–668 (1997). [Google Scholar]

PERMALINK

Three-dimensional aperture-driven photoacoustic tomography (3D-ADPAT)

Xuanhao Wang

Yang Xiao

Yuqi Wang

Xiali Gao

Yongdu Ruan

Xiaohui Yang

Yiyin Su

Dikui Zhou

Xiangdi Li

Yuqian Meng

Zhibo Xiao

Fan Meng

Ruofan Wang

Junhui Shi

Roles

Abstract

INTRODUCTION

RESULTS

Concept and system implementation of 3D-ADPAT

Fig. 1. Concept and system implementation of 3D-ADPAT.

PIFER for 3D-ADPAT

Fig. 2. Principle and simulation of the PIFER for 3D-ADPAT.

Performance evaluation and comparative analysis of 3D-ADPAT

Fig. 3. Phantom validation of 3D-ADPAT performance.

Fig. 4. Comparison of imaging performance between 3D-FS and 3D-ADPAT.

In vivo noninvasive anatomical and functional imaging with 3D-ADPAT

Fig. 5. In vivo noninvasive anatomical and functional imaging with 3D-ADPAT.

Array-based implementation of 3D-ADPAT

Fig. 6. Array-based implementation of 3D-ADPAT.

DISCUSSION

MATERIALS AND METHODS

3D-ADPAT system setup

Analysis of the acoustic field for a focused ultrasonic transducer

Numerical simulation setup

3D image reconstruction

Selection of ramp filter windows

Phantom and in vivo experiments

Acknowledgments

Funding:

Author contributions:

Competing interests:

Data, code, and materials availability:

Supplementary Materials

The PDF file includes:

Other Supplementary Material for this manuscript includes the following:

REFERENCES

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases