Robust automated three-dimensional speed-of-sound compensation for hemispherical photoacoustic computed tomography Authors: Xuran Zhou, Xiali Gao, Fan Meng, Hao Huang, Zhibo Xiao, Xuanhao Wang, Junhui Shi Published: 2026 ABSTRACT AbstractThree-dimensional hemispherical photoacoustic computed tomography provides vital structural information for non-invasive preclinical research. However, speed-of-sound heterogeneity between the coupling water and biological tissue induces wavefront aberrations, resulting in tangential blurring and ghosting artifacts. Traditional compensation typically relies on simplistic ellipsoidal fitting, which fails to capture irregular tissue boundaries, or utilizes complex multimodal ultrasound hardware that significantly increases system cost.To address these limitations, we introduce an automated, low-cost signal-domain sensing framework for a 1,024-element array that achieves high-fidelity three-dimensional boundary extraction without manual intervention or auxiliary hardware. By establishing a robust Median/Median Absolute Deviation based statistical noise baseline and a channel-specific, self-adaptive strong signal threshold, the framework maps the irregular acoustic interface directly from raw data, effectively isolating biological signals from intense scattering spikes and high-amplitude optical diffuser-induced artifact arcs. This signal-centric approach establishes an equivalent dual-speed-of-sound model through temporal compression, realigning distorted wavefronts with minimal computational overhead.Validated through three-dimensional simulations, irregular phantoms, and in vivo mouse imaging, the algorithm significantly restored the morphological fidelity of hepatic microvasculature. Furthermore, stress tests confirmed exceptional resilience even when the signal-to-noise ratio was artificially degraded to 30.9 dB. This robust architecture enables high-fidelity three-dimensional imaging in acoustically challenging and high-interference biological environments.Keywords: Photoacoustic, PACT, Speed-of-Sound Heterogeneity, Automated Boundary Segmentation Introduction Combining the high sensitivity of optical absorption with the superior penetration of ultrasound, photoacoustic imaging (PAI) has emerged as a cornerstone of modern biomedical diagnostics [1], [2], [3], [4], [5], [6]. The technique captures photoacoustic (PA) signals—ultrasonic waves emitted by tissues upon laser excitation—to map the optical properties of deep-seated structures [7], [8]. This synergy allows photoacoustic tomography (PAT) to achieve high-resolution molecular and functional imaging, even at centimeter-scale depths, making it invaluable for oncology, neurology, and hemodynamics [9], [10], [11], [12], [13], [14]. However, the fidelity of reconstruction in PAT remains challenging; gold-standard algorithms like delay-and-sum (DAS) [15] and back-projection typically assume a homogeneous speed of sound (SoS), a simplification that often overlooks the complex acoustic environment of real biological tissues. Despite its high-contrast capabilities, the practical fidelity of PAT is severely compromised by the intrinsic SoS heterogeneity of biological environments [16], [17], [18]. While minor variations exist within internal tissues, the most significant acoustic mismatch occurs at the interface between the coupling medium and the biological target. Because the PA signals must traverse long propagation paths before reaching the transducers, even a subtle SoS discrepancy accumulates into substantial phase distortions and time-of-flight errors [19], [20], [21], [22]. In recent years, researchers have attempted to mitigate this through dual-SoS reconstruction. Traditional approaches often rely on image-domain segmentation, such as fitting elliptical boundaries to the subject's contour [23]. However, these methods are primarily optimized for 2D geometries and fail to accommodate the complex, non-parametric surfaces of real 3D anatomy.While pioneering studies have demonstrated that a heuristically fitted uniform speed-of-sound can efficiently improve image quality and is highly practical for many optoacoustic setups [24], this simplification presents limitations for large-scale 3D hemispherical arrays. The primary difference from 2D systems is the vastly increased number of sensing channels. This dimensionality bottleneck, combined with the core research difficulty of highly irregular tissue boundaries, exacerbates out-of-plane wavefront aberrations. A single averaged SoS cannot fully compensate for the localized temporal shifts across 1024 elements, inevitably leading to spatial delocalization of deep-seated absorbers. Furthermore, while multimodal systems integrating ultrasound (US) have been proposed to assist in boundary identification [25], [26], [27], the hardware complexity and high computational cost of US-PA co-registration hinder their widespread adoption in high-speed applications. A recent breakthrough in 2D ring-array PAT introduced a signal-domain SoS correction method that identifies the water-tissue interface through a variance-based analysis of the PA sinogram [28]. However, our investigation reveals that this 2D-centric approach is critically incompatible with 3D hemispherical PACT systems. First, unlike the full-enclosure geometry of a 360° ring array, a hemispherical configuration inherently suffers from a limited detection aperture, leading to substantial signal loss and incomplete data coverage—a "geometric missing-wedge" problem that complicates 3D boundary sensing. Second, the integration of optical diffusers in high-density 3D systems generates intense interference signals (background clutter) within the sinogram, which causes traditional variance-based segmentation to fail [29]. More than that, unlike 2D ring-array systems that utilize physical cylindrical focusing to concentrate acoustic energy within a specific slice, 3D hemispherical arrays must employ non-focused elements with a wide acceptance angle to ensure full volumetric coverage, inherently reducing the single-channel raw signal-to-noise ratio (SNR). Mathematically, the 2D ring-array method relies on minimizing the sum of log-variances (typically via the Akaike Information Criterion) to partition the signal and detect transition points. Because variance calculations rely on squared deviations from the mean, they are inherently hypersensitive to large-amplitude outliers. In a 3D hemispherical system, the intense, non-biological scattering spikes generated by the wide-field optical diffuser act as massive statistical outliers. These high-amplitude interference arcs completely dominate the variance summation, causing traditional algorithms to incorrectly lock onto the optical artifacts rather than the true biological tissue boundaries. There is a critical need for a framework that can maintain high robustness in these high-dimensional, low-SNR environments. Specifically, an automated mechanism is required that can sense the 3D boundary with minimal computational overhead—transitioning from manual, operator-dependent image-domain segmentation to a signal-domain compression strategy that remains stable even under the sophisticated geometric constraints of hemispherical acoustic sensing. To resolve this, we have developed a more robust framework using an adaptive noise floor algorithm to accurately isolate first-arrival time indices. This method not only overcomes the SNR challenges inherent in 3D clinical imaging but also equivalently maps the 3D acoustic boundary with minimal computational cost. By addressing these 3D-specific depth aberrations and complex out-of-plane distortions that 2D models cannot resolve, our approach ensures high-throughput, automated SoS compensation with superior robustness against the sophisticated acoustic environments of hemispherical arrays. The efficacy of this framework is validated through a rigorous series of 3D numerical simulations, physical phantom experiments, and in vivo mouse vascular imaging. Our results demonstrate that this automated sensing architecture can rectify complex wavefront aberrations with exceptional precision, achieving a mean radial error (MRE) of approximately 0.124 mm. More importantly, by providing a high-robustness solution for 3D acoustic boundary mapping, this work solves the long-standing bottleneck of automated SoS correction in sophisticated 3D PACT systems. This transition from manual calibration to automated, signal-centric compensation paves the way for high-throughput preclinical screening and reliable hemodynamic monitoring in acoustically heterogeneous environments. Methods Dual-SoS in PACT In a non-viscous and thermally confined biological medium, the generation and subsequent propagation of PA waves—triggered by the absorption of short-pulse laser energy—can be characterized by the following inhomogeneous wave equation [30], [31]: In this expression, p(r,t) represents the induced acoustic pressure at spatial coordinate r and time t, while v denotes the local SoS. The right-hand side describes the source term, where H(r,t) is the heating function (energy deposited per unit volume and time), β is the isobaric volume expansion coefficient, and Cp is the specific heat capacity at constant pressure.Under the assumption of impulsive excitation (where the laser pulse width is negligible relative to the acoustic relaxation time), the solution for the forward propagation of the pressure field can be derived using the Green’s function method: where p0(r′) represents the initial photoacoustic pressure at the source. This integral solution explicitly demonstrates that the detected pressure p(r,t) is intrinsically coupled to the SoS, v, through the time-of-flight (TOF) relationship. Consequently, the spatial fidelity of image reconstruction is strictly contingent upon the accuracy of the SoS profile utilized in the back-projection or inversion process.In practical imaging scenarios, a significant discrepancy often exists between the actual acoustic environment—referred to as the Propagation Model (PM)—and the simplified velocity distribution assumed during computation, known as the Reconstruction Model (RM). A standard RM typically assumes a homogeneous SoS, which fails to account for the acoustic refraction and phase delays inherent in the heterogeneous PM. While advanced dual-SoS compensation methods can rectify these artifacts, most existing solutions—including those leveraging deep learning architectures or auxiliary ultrasound hardware—require prohibitive computational resources or hardware modifications. Therefore, developing a strategy that achieves dual-SoS fidelity while maintaining the computational efficiency of a single-SoS reconstruction model is essential for high-throughput 3D PACT applications. Overall framework of automated SoS The conceptual framework of the proposed automated SoS compensation for 3D hemispherical PACT is detailed in Fig. 1a. The imaging sequence is initiated by 1064 nm laser pulses 10 ns pulse width, 50 Hz repetition rate), which are uniformly projected through an engineered optical diffuser to provide wide-field excitation of the target chromophores. Upon absorption, the resulting thermoelastic expansion generates broadband PA waves whose spectral characteristics are primarily determined by the morphological scales of the tissue. In an ideal homogeneous medium, these waves are assumed to propagate as spherical or cylindrical fronts; however, in a realistic clinical setting, these acoustic wavefronts must traverse the interface between the biological tissue (SoS V2) and the acoustic coupling water (SoS V1). Because the wave frequency remains invariant across media while the wavelength adapts according to λ=c/f, this velocity mismatch leads to severe phase decoherence and cumulative temporal shifts. In a high-density 3D volumetric geometry, these aberrations manifest as complex out-of-plane distortions and "ghosting" artifacts that are substantially more difficult to rectify than the in-plane errors encountered in traditional 2D ring-array configurations, where the refractive paths are largely restricted to a single detection plane.To capture these high-dimensional wavefields with high spatial fidelity, we utilize a custom-built, high-density 1,024-element hemispherical ultrasonic transducer array (HUTA) integrated into a 20 cm diameter shell. The transducers are distributed via a Fibonacci spiral algorithm to achieve near-uniform spatial sampling and minimize side-lobe artifacts across the 3D surface. Synchronous data acquisition at a 40 MHz sampling rate generates a raw 2D PA sinogram, a high-dimensional data matrix where each row represents the temporal pressure response of a single sensing element. In such a large-scale sensing architecture, traditional SoS correction methods—specifically those relying on manual boundary identification or iterative image-domain segmentation—face a computational scaling bottleneck. Manual segmentation across thousands of channels is practically impossible, while iterative algorithms suffer from excessive computational overhead, rendering them incompatible with real-time hemodynamic monitoring.The core of our methodology addresses this bottleneck by transforming the distorted 3D wavefield into a unified reference domain through a non-iterative sensing logic. By identifying the Transition Point (TP) directly within the raw sinogram, the framework effectively partitions each A-line into distinct dual-SoS intervals. To achieve equivalent SoS compensation without the computational cost of ray-tracing or complex iterative refractive modeling, we define a time-mapping operator utilizing an indicator function I(⋅), which represents the corrected effective pressure signal Peff(τ) as a single-line analytical mapping of the original acquisition: Conceptual framework of the automated signal-domain SoS compensation and the adaptive boundary-sensing algorithm. (a) Schematic of the 3D hemispherical PACT system and the equivalent dual-SoS correction logic. Photoacoustic (PA) waves propagate through a heterogeneous medium (V2 for tissue, V1 for water) and are captured by a 1,024-element hemispherical transducer array (HUTA). The correction process involves identifying the Transition Point (TP) in the raw sinogram and applying a signal-domain compression strategy to realign distorted wavefronts into an equivalent single-SoS domain. (b) Detailed flowchart of the multi-layered, adaptive segmentation algorithm. The framework utilizes a Hilbert-transformed envelope, a statistical noise baseline (μ0,σ0), and a dual-verification mechanism (Strong Signal Ratio and Continuous Segment Length) to achieve robust boundary extraction even under low SINR conditions. A fail-safe fallback logic is integrated to ensure high-precision segmentation in cases of extreme signal degradation. Within this mapping, β denotes the amplitude scaling factor—justifiably assumed as a unit factor given that amplitude variations over short acoustic delays are negligible—while η represents the dimensionless SoS ratio V1/V2 and τTP corresponds to the automatically sensed temporal coordinate of the tissue-water interface. By implementing this automated compression across all 1024 channels, the system establishes an equivalent dual-SoS model that realigns the distorted wavefronts into a uniform domain prior to reconstruction. This signal-centric transformation enables the use of standard, high-speed reconstruction parameters while restoring sub-millimeter structural sharpness, effectively solving the trade-off between computational efficiency and imaging fidelity in acoustically heterogeneous 3D environments. To ensure optimal compensation under varying experimental conditions (e.g., localized temperature gradients or specific tissue compositions during in vivo imaging), the specific values for the outer coupling medium (V1) and the inner biological tissue (V2) are determined through a step-wise, image-sharpness-based parameter sweeping strategy prior to the final 3D volumetric reconstruction. First, within a physically reasonable acoustic range, target slices are iteratively reconstructed using candidate V1 values. The optimal V1 is established by selecting the value that yields the sharpest and most highly contrasted reconstruction of the superficial microvasculature located immediately adjacent to the outer contour. Once V1 is fixed, a subsequent parameter sweep is performed for the inner region. The optimal V2 is determined by identifying the value that maximizes the Contrast-to-Noise Ratio (CNR) and minimizes ghosting artifacts for the deep-seated internal tissue structures. This localized approach allows the signal-domain compression framework to adaptively accommodate the slight dynamic acoustic variations inherent in live animal imaging, ensuring maximum spatial fidelity across the entire volume. Adaptive signal-domain boundary segmentation and robust TP extraction To achieve the non-iterative dual-SoS compensation described in the preceding section, the precise localization of the TP—representing the first acoustic arrival at the tissue-water interface—is paramount. In high-density hemispherical transducer arrays, this task is fundamentally challenged by diffuser-induced clutter. High-amplitude scattering signals from the optical diffuser are ubiquitously superimposed onto the raw 3D sinograms, often masking the true acoustic wavefront. To resolve these persistent interferences, we developed a multi-layered, adaptive segmentation algorithm designed to maintain high-fidelity boundary extraction even under low signal-to-interference-plus-noise ratio (SINR) conditions, as illustrated in the procedural workflow of Fig. 1b.The sensing logic begins with preprocessing the raw A-line through a bandpass filter and analytic envelope extraction via the Hilbert transform. Existing signal-domain approaches typically identify the boundary by searching for a temporal index k that minimizes the combined statistical variance of the pre- and post-boundary segments ( klog(σ1,k2)+(N−k)log(σk+1,N2)). However, under the low-SINR conditions of hemispherical PACT, the large-amplitude optical scattering artifacts cause the squared deviations in variance calculations to exponentially explode, leading to inevitable false segmentations. To overcome this fundamental mathematical limitation, our algorithm abandons global variance minimization. Instead, we developed a robust statistical framework based on the Median and Median Absolute Deviation (MAD) of the noise baseline. Unlike variance, which evaluates the L2-norm and is easily skewed by extreme values, MAD is intrinsically outlier-resilient with a high breakdown point (evaluating the L1-norm). This approach provides a stable estimation of the noise floor (μ0) and the effective standard deviation (σ0), establishing a dynamic reference that cannot be corrupted by diffuser-induced interference spikes. A distinctive innovation of this framework is the unsupervised calibration of the strong signal threshold (strong_thr), which is adaptively adjusted for every channel based on the peak envelope intensity within the localized search segment. This self-adaptive thresholding ensures the system remains sensitive to genuine biological signals while preserving high anti-interference capability across varying imaging depths and signal strengths inherent in 3D hemispherical sensing.To further fortify the detection integrity against the intense scattering spikes visible in the 3D sinograms, we integrated a dual-verification safeguard mechanism. Once a candidate rising edge is identified by the adaptive low threshold, it must pass through two sequential logical gates:Module 1 (Strong Signal Ratio): Evaluates the local energy density (K) to filter out non-coherent, stochastic electronic noise.Module 2 (Continuous Segment Length): Mandates a minimum run-length (Max_Runlen) to distinguish the sustained acoustic wavefront of the biological target from transient, narrow-band scattering spikes.To ensure total algorithmic robustness in extreme cases of signal degradation or significant signal loss—common in the limited-aperture geometry of hemispherical arrays—a fail-safe fallback mechanism is integrated. If the dual-verification criteria (AND logic) are not satisfied by any candidate, the algorithm automatically reverts to a gradient-based search, identifying the point of maximum derivative from the smoothed envelope. This multi-staged strategy ensures that the boundary segmentation remains both high-precision and artifact-resilient, providing a stable acoustic prior for the subsequent signal-domain compression and high-resolution 3D PACT reconstruction. To quantitatively evaluate the geometric fidelity of the extracted boundary against the actual tissue interface, the Mean Radial Error (MRE) was calculated. The MRE is defined as: where N represents the total number of effective transducer channels, and r denotes the radial distance from the coordinate center to the mapped acoustic boundary.Furthermore, transitioning to a high-density 1,024-element hemispherical array makes the system highly sensitive to algorithmic time complexity. Traditional variance-based segmentation requires an exhaustive traversal of all possible transition indices, yielding a time complexity of O(L2) per channel (where L is the A-line signal length). In contrast, our Median/MAD-based dual-verification logic operates via a single-pass linear scan. This inherently reduces the time complexity to O(L) per channel, dropping an entire order of magnitude and effectively resolving the computational bottleneck of 3D boundary sensing. Results Simulation experiment To rigorously evaluate the proposed 3D SoS correction algorithm, a biologically inspired 3D vascular phantom was developed within a 256×256×256 computational grid with an isotropic voxel size of 312.5 μm. As illustrated in Fig. 2a, the phantom features a complex 3D vascular tree generated through a multi-step morphological process. Initially, a 2D vascular skeleton was extracted and projected into 3D space, where stochastic depth fluctuations were introduced using a trigonometric modulation: zoffset=25⋅sin(x/30)⋅cos(y/25), ensuring a realistic 3D volumetric distribution. To mimic physiological microvasculature, a dynamic radial profile was applied, with vessel radii tapering from 1.1 mm (3.5 pixels) at the main trunk to 0.375 mm (1.2 pixels) at the distal capillaries. The heterogeneous acoustic environment was modeled as a transverse egg-shaped "inner" tissue region (cinner = 1490 m/s) embedded within an "outer" coupling medium (couter = 1580 m/s). Unlike standard spherical models, the SoS boundary was defined by a distorted ellipsoid equation to test geometric robustness: where a= 30 mm, b= 38 mm, and k = 0.25 is the distortion coefficient creating non-parametric asymmetry along the x-axis.Photoacoustic (PA) signals were acquired using a virtual hemispherical transducer array with a radius of 35 mm (Fig. 2a, blue shell), populated by 1024 point-like elements distributed via a Fibonacci spiral algorithm. This high-density configuration is essential for capturing the complex 3D wavefield. Wave propagation was simulated using the k-Wave pseudospectral solver with the temporal sampling interval (dt) determined by the CFL stability criterion. A core innovation of this work is the automated identification of the TP directly within the raw sinogram to facilitate "equivalent" dual-SoS processing. As visualized in Fig. 2b, the raw sinogram (top) contains significant temporal shifts induced by the slower cinner. For each transducer channel, our algorithm utilizes an adaptive noise floor to detect the first-arrival time index (tidx) of the PA signals, which corresponds to the water-tissue interface.Once the TP is localized, an equivalent dual-SoS treatment is performed by applying a signal-domain compression to the latter portion of the A-line. This "shortcut" avoids the heavy computational burden of iterative image-domain segmentation by realigning the distorted wavefield into an equivalent homogeneous domain (Fig. 2b, bottom zoom-in). To verify the segmentation accuracy without significant extra computation, the identified TPs were mapped into the image domain by assigning a peak magnitude (106) to the corresponding time indices. The resulting visualization (Fig. 2d) demonstrates that our signal-domain algorithm precisely reconstructs the 3D SoS boundary across all anatomical planes.The final reconstruction performance was assessed by comparing the Ground Truth (Fig. 2c) against uncompensated and compensated results. Under a standard single-SoS assumption (1500 m/s, Fig. 2e), the reconstruction suffers from severe tangential blurring and spatial delocalization, particularly in deeper tissue regions. The mismatch leads to incoherent back-projection, manifesting as "ghosting" artifacts and a loss of structural integrity. In contrast, the proposed adaptive reconstruction (Fig. 2f) successfully rectifies these wavefront aberrations. The spatial fidelity of the complex 3D bifurcations and dynamic vessel diameters is markedly restored. This demonstrates that our signal-domain compression approach provides a robust, low-cost solution for high-fidelity 3D PACT imaging in acoustically heterogeneous media. Numerical validation using a biologically inspired 3D vascular phantom in a transverse heterogeneous medium. (a) 3D simulation setup featuring a vascular tree with dynamic radii (0.375–1.1 mm) and stochastic depth fluctuations embedded in a non-parametric, egg-shaped SoS field (1490 m/s inner, 1580 m/s outer). (b) Comparison between the raw (top) and corrected (bottom) sinograms; the zoom-in view highlights the successful realignment of PA signals following signal-domain compression. (c) Ground Truth (GT) morphology across three anatomical planes. (d) Visualization of the automated boundary sensing, where TPs are mapped into the image domain with a peak intensity (106) to verify spatial accuracy. (e) Single-SoS reconstruction (1500 m/s) exhibiting severe tangential blurring and ghosting artifacts. (f) Dual-SoS reconstruction using the proposed framework, demonstrating restored structural fidelity and sub-millimeter resolution across all planes.(Scale bar: 10 mm). Experimental validation and in vivo performance To bridge the gap between numerical theory and physical application, the proposed framework was implemented on a custom-built 3D PACT system for experimental validation. The hardware architecture, illustrated in Fig. 3a, centers on a 20-cm diameter hemispherical shell integrated with 1024 Fibonacci-arranged piezoelectric transducers (3.15 MHz center frequency). An Nd:YAG laser providing 1064 nm pulses is coupled with a beam diffuser to deliver uniform illumination across a 4 cm spot. Following system characterization, a polygonal phantom composed of agarose and 0.02% India ink was imaged to assess the algorithm’s performance in sensing non-parametric, sharp-edged geometries (Fig. 3b). While traditional single-SoS reconstruction typically produces severe refractive blurring and geometric warping at linear facets and vertices, our sensing logic accurately delineated the phantom’s complex polygonal profile, demonstrating high geometric fidelity even at the sharpest vertices. The translational potential of the method was further evaluated through in vivo imaging of a mouse liver across three representative anatomical sections (Fig. 3c). These images reveal the intricate vascular architecture of the hepatic lobes, where the high-density array allows for deep-tissue penetration and the capture of fine structural details that are often obscured by acoustic heterogeneity. To visualize the boundary sensing accuracy in this complex biological environment, the TPs identified across all channels were assigned a peak intensity value (106) and projected into the image domain. This mapping, manifested as a hollow ring in the rightmost column of Fig. 3c, shows an exceptional fit between the sensed boundary (green dashed circles) and the actual anatomical contours of the liver. It is worth noting that due to the highly irregular 3D boundaries of the biological tissue, a full-volume Z-axis MIP (comprising 600 voxels) inevitably introduces a geometric projection effect, where structures from adjacent planes visually overlap the 2D contour. To strictly decouple this effect and demonstrate the exact layer-specific fit of our algorithm, thin-slab reconstructions (with a reduced Z-depth of 30 voxels) are provided in Supplementary Note 1 and Fig. S1. The precisely mapped interface confirms that the algorithm can distinguish the tissue boundary from the surrounding coupling medium with sub-millimeter reliability across varying anatomical depths. The robustness of the signal-domain identification is further scrutinized through the sinogram analysis presented in the bottom row of Fig. 3. A significant challenge in high-density 3D hemispherical systems is the intense background interference—visible as a prominent high-amplitude arc at the leading edge of the raw sinogram—caused by the optical diffuser. In such high-clutter environments, conventional 2D signal-domain algorithms, such as those relying on the minimum sum of variances, frequently fail as they are unable to distinguish between the interference arc and the true tissue-coupling medium interface. In contrast, our adaptive noise floor algorithm remains stable, effectively isolating the first-arrival biological signals from the diffuser-induced clutter. The resulting "Green TP band" represents the ensemble of transition points collected from all 1024 transducer channels. Each green pixel within these bands identifies the precise phase transition for an individual channel, tracking the irregular mouse body contour with high fidelity. As shown in the magnified sinogram views, these TP coordinates remain strictly adherent to the biological signal despite the dense interference background. By applying signal-domain compression based on these accurate prior coordinates, the dual-SoS corrected sinograms exhibit a marked realignment of wavefronts, effectively compensating for the temporal shifts induced by the varying speed of sound. This automated, non-iterative sensing mechanism ensures that the 3D PACT system can provide high-resolution, artifact-free imaging of complex biological structures without the need for manual boundary calibration or high-cost multimodal hardware assistance. Experimental validation and in vivo mouse imaging demonstrating interference immunity. (a) Custom-built 1,024-channel 3D PACT system instrumentation. (b) Geometric fidelity test using a polygonal agarose phantom (0.02% India ink); the algorithm accurately captures sharp refractive boundaries. (c) In vivo mouse liver imaging across three representative sections. The top row displays full-volume Z-axis Maximum Intensity Projections (MIPs) of the 3D vascular networks, illustrating the high-precision fit between the mapped boundary (hollow rings, representing the transition points for one specific 2D Z-slice) and actual anatomical contours (green dashed circles). The bottom panels elucidate the signal-domain mechanism: the "Green TP bands" track the physical tissue-water interface with high stability, successfully isolating biological signals from the intense, leading-edge interference arc caused by the optical diffuser. Comparative analysis of reconstruction performance The quantitative superiority and spatial fidelity of the proposed dual-speed correction framework are further substantiated through a localized ROI analysis of the hepatic microvasculature (Fig. 4). A side-by-side comparison between the dual-SoS compensated images (Fig. 4a) and the standard single-SoS reconstructions (Fig. 4b) reveals the profound impact of acoustic heterogeneity on deep-tissue 3D PACT. In the uncorrected images, the mismatch between the coupling water and biological tissue induces severe tangential blurring and phase decoherence, causing the fine vascular networks to appear as diffused, low-contrast structures. To explicitly demonstrate the depth-dependent accumulation of these aberrations, a direct comparison between superficial (e.g., abdominal aorta) and deep-seated hepatic regions is provided in Supplementary Note 2 and Fig. S2, confirming that dual-SoS correction is absolutely essential for deep-tissue recovery. In contrast, our automated sensing-based correction effectively realigns the distorted wavefronts, restoring the morphological integrity of the 3D vascular tree with significantly enhanced boundary sharpness and structural continuity. Quantitative ROI analysis of hepatic microvasculature for resolution and contrast enhancement. (a) Dual-SoS compensated images showing clear vascular continuity. (b) Standard single-SoS reconstructions suffering from refractive blurring. (c) Magnified views of the regions of interest (ROI); the dual-SoS approach (blue boxes) resolves fine bifurcations obscured in the single-SoS results (yellow boxes). (d) Normalized PA amplitude profiles along the red dashed lines in (c). The significant narrowing of the intensity profile and the improved Contrast-to-Noise Ratio (CNR) in the dual-SoS curves (blue) quantitatively confirm the restoration of sub-millimeter structural sharpness. The line profiles across representative vessels, indicated by the red dashed lines in Fig. 4c, provide a direct quantitative measure of the resolution enhancement. As plotted in the rightmost column of Fig. 4, the normalized PA amplitude profiles demonstrate a marked narrowing of the intensity profile for the dual-SoS method (blue curves) compared to the single-SoS results (yellow curves). The peaks in the dual-SoS profiles are not only sharper but also exhibit significantly higher signal-to-background contrast. Specifically, fine vascular features that are almost entirely buried within the side-lobes and noise in the single-SoS reconstruction are clearly resolved as distinct, high-amplitude peaks following our signal-domain compression. This improvement in the Contrast-to-Noise Ratio (CNR) and spatial resolution is a direct consequence of the constructive interference enabled by our adaptive boundary-sensing algorithm. To rigorously quantify these enhancements, we extracted the specific metrics from the representative deep-tissue vascular profile (Fig. 4d). In the uncompensated single-SoS reconstruction, the accumulated aberrations severely degraded the signal, yielding a low CNR of 11.7 and a broadened spatial resolution (FWHM) of 0.49 mm. In stark contrast, the automated dual-SoS compensation effectively re-focused the wavefronts, dramatically increasing the CNR to 18.5 and sharpening the FWHM to 0.38 mm. By accurately accounting for the 3D refractive interface, the system achieves the sub-millimeter precision necessary for reliable whole-body imaging, proving that our non-iterative signal-centric architecture can maintain high-fidelity performance even in the most acoustically challenging biological environments. To comprehensively evaluate the advantages of our proposed framework, its performance was compared against two established SoS correction approaches: the ellipsoidal boundary fitting method and the variance-based signal-domain correction method (detailed in Supplementary Note 3 and Fig. S3). While the ellipsoidal fitting method is highly sensitive to local morphological variations, leading to spatially inconsistent correction, and the variance-based method suffers from severe interference artifacts in hemispherical systems, our proposed automated framework demonstrates superior and spatially robust structural restoration. Quantitatively, our method achieved the highest Contrast-to-Noise Ratio (CNR) and the sharpest spatial resolution (measured by FWHM) among the compared techniques. Beyond spatial fidelity, the algorithmic efficiency was quantitatively validated to support our claims of low computational overhead. Tested on a standard workstation ( Intel Core i7–11700k / NVIDIA RTX 3090), the traditional O(L2) variance-summation approach required approximately 15.2 s to segment a full 1,024-channel volumetric frame. In stark contrast, our O(L) adaptive framework completed the identical 3D segmentation task in just 0.12 s. This massive reduction in runtime strictly substantiates the high-throughput capability of our method, rendering it highly compatible with real-time in vivo monitoring applications. Noise immunity and reliability assessment To evaluate the algorithm’s resilience in challenging clinical scenarios where SNR are often compromised by acoustic attenuation and deep-tissue scattering, we conducted a rigorous stress test by artificially degrading the signal-to-noise ratio (SNR) of the raw sinograms (Fig. 5). The SNR of the sensing channel was explicitly defined based on the peak signal amplitude relative to the root-mean-square (RMS) of the background noise: Stress test of algorithmic resilience across varying stochastic noise levels. (a-c) Boundary sensing and 3D reconstruction performance under baseline (37.6 dB), moderately degraded (34.8 dB), and significantly degraded (30.9 dB) SNR conditions. The TP bands remain strictly adherent to the tissue interface despite the severe degradation of the sinogram’s SNR. (d) Line profile overlap across the three noise gradients, demonstrating the morphological stability of the proposed framework. (e) Box plot of sample offsets across the 1,024-element array; the tightly bounded interquartile ranges and lack of erratic outliers confirm the statistical robustness of the adaptive noise floor algorithm in complex clinical sensing scenarios. Three representative sensing environments were established to observe the performance threshold of the adaptive boundary-sensing logic: a baseline state with an inherently high SNR (37.6 dB), a moderately degraded SNR state (34.8 dB), and a significantly degraded SNR state (30.9 dB). Qualitatively, the boundary sensing performance remains remarkably consistent despite the visible degradation of the sinogram’s signal-to-background contrast (Fig. 5a–c). Even in the 30.9 dB case, where the raw PA signals are nearly submerged within the noise floor, the extracted transition points (TP bands) exhibit a negligible deviation from the baseline. This stability stems from the adaptive noise floor algorithm’s ability to distinguish coherent first-arrival biological signals from non-coherent stochastic fluctuations, a task where traditional variance-based 2D algorithms typically fail due to their reliance on local intensity thresholds.Quantitatively, the high degree of morphological preservation is reflected in the line profiles across the hepatic vessels (Fig. 5d). The normalized PA amplitude curves for all three noise gradients show exceptional overlap, particularly in the peak positions and structural widths, indicating that the dual-SoS compensation maintains its spatial fidelity even as the signal environment deteriorates. To provide a comprehensive statistical evaluation of the 3D sensing logic, the distribution of sample offsets across the entire 1,024-element transducer array was analyzed (Fig. 5e). This box plot illustrates the localized boundary-sensing shifts for every individual channel. While the median sample offset naturally increases as the noise power effectively doubles—shifting from the green baseline to the blue extreme case—the interquartile ranges remain tightly bounded. The absence of erratic outliers, even at 30.9 dB, confirms that the algorithm preserves a sub-millimeter precision manifold across the high-density array. This "insensitivity" to high noise levels underscores the system's reliability for high-throughput preclinical screening, ensuring that the detected acoustic priors remain an accurate reflection of the physical tissue-coupling medium interface regardless of sensing conditions. Conclusion and discussion In this work, we have presented and validated an automated, low-cost signal-domain sensing framework designed to achieve high-fidelity 3D-SoS compensation in PACT. Inherent velocity heterogeneity between the acoustic coupling medium and biological tissue has long introduced wavefront aberrations that compromise deep-tissue imaging, leading to severe tangential blurring and ghosting artifacts. Traditional compensation frameworks typically depend on simplistic ellipsoidal fitting, which fails to capture irregular anatomical boundaries, or rely on complex multimodal ultrasound hardware that significantly increases system costs. Furthermore, conventional manual, operator-dependent image-domain segmentation is highly inefficient and ill-suited for high-throughput applications. Our signal-centric approach effectively addresses these limitations by establishing an equivalent dual-SoS model through temporal compression, realigning distorted wavefronts across a high-density, 1,024-element hemispherical transducer array with minimal computational overhead and without requiring auxiliary hardware integrations. A key innovation of this framework lies in its capacity to navigate the sophisticated geometric constraints and complex out-of-plane distortions inherent to 3D hemispherical acoustic sensing, which standard 2D models cannot resolve. Unlike 2D ring-array systems that utilize physical cylindrical focusing to concentrate acoustic energy within a specific slice, 3D hemispherical arrays must employ non-focused elements with a wide acceptance angle to ensure full volumetric coverage, inherently reducing the single-channel raw SNR. To resolve this dimensionality and noise bottleneck, our framework discards fragile fixed-threshold techniques in favor of a robust Median/MAD-based statistical noise baseline combined with a channel-specific, self-adaptive strong signal threshold. To further fortify detection integrity against intense scattering spikes and prominent system artifacts—such as the high-amplitude optical diffuser-induced artifact arc—we integrated a dual-verification safeguard mechanism based on local energy density and continuous segment length, backed by a gradient-based fail-safe fallback engine. This multi-staged, signal-domain strategy guarantees high-precision and outlier-resilient first-arrival wavefront mapping, ensuring that the extracted 3D acoustic boundary remains strictly adherent to the true irregular tissue interface even under severe signal degradation. The physical performance and quantitative precision of this automated architecture were rigorously substantiated across numerical, phantom, and in vivo sensing environments. Validation using irregular tissue phantoms confirmed that the boundary-sensing logic successfully adapts to non-parametric geometries that cause traditional single-SoS reconstructions to fail. In in vivo mouse liver imaging across multiple anatomical cross-sections, the framework achieved an exceptional MRE of approximately 0.124 mm. By enabling the constructive interference of back-projected signals via accurate signal-domain mapping, we significantly restoring morphological fidelity and enhancing the contrast-to-noise ratio (CNR). Crucially, stress tests under simulated conditions confirmed that the framework maintains high structural preservation and tight statistical consistency across all 1024 channels even when the SNR was degraded to 30.9 dB. Despite these substantial performance gains, certain boundary conditions remain to be addressed in future iterations. The current equivalent dual-SoS compression model assumes a relatively sharp, well-defined acoustic interface between two primary media. While this assumption is highly effective for rodent imaging and whole-body preclinical screening where the skin-water transition represents the dominant refractive boundary, more complex clinical scenarios involving multi-layered internal heterogeneities—such as bone-tissue interfaces within the human skull—will necessitate the extension of this logic toward multi-layered or gradient-based SoS profiles. Additionally, future efforts will focus on migrating the current signal-domain architecture into real-time, GPU-accelerated reconstruction engines to facilitate live, motion-corrected 3D photoacoustic videos. Integrating the adaptive tracking mechanism with machine learning-based semantic segmentation could also further refine boundary identification in low-contrast regions with overlapping biological signatures. In conclusion, the proposed automated first-arrival wavefront mapping architecture provides a scalable, robust, and high-precision solution for SoS compensation in sophisticated 3D PACT systems. By successfully bridging the gap between automated signal-domain sensing and 3D volumetric image fidelity, this framework effectively eliminates the long-standing trade-off between hardware complexity and reconstruction accuracy. This transition from manual calibration to low-cost, automated, and artifact-resilient compensation establishes a reliable foundation for high-throughput preclinical screening and reliable diagnostic hemodynamic monitoring in acoustically challenging biological environments. Funding This work was supported by the National Key Research and Development Program of China (No. 2023YFC2410900); State Key Laboratory of Ocean Sensing Operation Funding (No. 4400-4CZ260100); the National Key R&D Program of China (No. 2023YFB3211200). CRediT authorship contribution statement Xuran Zhou: Writing – review & editing, Writing – original draft, Software, Project administration, Investigation, Formal analysis, Data curation, Conceptualization. Xiali Gao: Methodology. Fan Meng: Data curation. Hao Huang: Software. Zhibo Xiao: Data curation. Xuanhao Wang: Supervision. Junhui Shi: Writing – review & editing, Supervision. Declaration of Generative AI and AI-assisted technologies in the writing process During the preparation of this manuscript, the authors used Gemini 3.1 Pro to assist with code development for statistical analyses and figure generation, as well as language editing of the manuscript text. All AI-assisted outputs were reviewed, verified against primary data, and edited by the authors. The authors take full responsibility for the scientific content, interpretation, and conclusions of the publication. No generative AI tools were used to generate primary data, design experiments, or draw scientific conclusions. Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Biographies Xuran Zhou received his bachelor's degree in Optoelectronic Information Science and Engineering from Nanjing University of Posts and Telecommunications and his master's degree in Optical Engineering from Jiangnan University. He is currently pursuing a Ph.D. in Biomedical Engineering at Zhejiang University. His current research focuses on photoacoustic imaging algorithms. Xiali Gao earned his bachelor’s and master’s degrees in optical engineering from Jiangnan University and is currently pursuing a Ph.D. in biomedical engineering at Zhejiang University. His research focuses on optical imaging, photoacoustic imaging, and image enhancement. Fan Meng is a doctoral student majoring in biomedical engineering at Zhejiang University. she graduated from China Pharmaceutical University. Her main research direction is the biomedical application of photoacoustic imaging. Hao Huang received his bachelor's degree in Automation from Harbin Engineering University and his master's degree in Control Science and Engineering from Harbin Institute of Technology. He is currently pursuing a Ph.D. in Biomedical Engineering at Zhejiang University. His research currently focuses on ultrasound and photoacoustic multi-modality imaging. Zhibo Xiao earned a Bachelor of Science in Software Engineering from China University of Petroleum and a Master of Science in Electronic Information from Jiangnan University. He is currently pursuing a Ph.D. in Biomedical Engineering at Zhejiang University. His research focuses on the integration of artificial intelligence and biomedical imaging. Xuanhao Wang is currently an Associate Researcher at Zhejiang Lab. He received his bachelor's degree from the School of Optics and Photonics at the Beijing Institute of Technology in 2017, and earned his Ph.D. from the Biophotonics Laboratory in the Department of Electronic Engineering at Tsinghua University. His research primarily focuses on developing advanced photoacoustic computed tomography systems and exploring their pathbreaking applications in addressing biological problems. Junhui Shi received his B.Sc. in chemical physics from the University of Science and Technology of China. Then, he continued to study chemistry and received his Ph.D. at Princeton University. He was working on theoretical chemical dynamics and experimental nuclear magnetic resonance spectroscopy. Afterward, he worked on photoacoustic imaging in the Department of Biomedical Engineering of Washington University in St. Louis, Missouri, and then in the Department of Medical Engineering of Caltech in Pasadena, California. Currently, he is a senior researcher at Zhejiang Lab, Hangzhou, China. His research interests include photoacoustic microscopy, photoacoustic tomography, nuclear magnetic imaging and photoacoustic spectroscopy. Footnotes Supplementary data associated with this article can be found in the online version at doi:10.1016/j.pacs.2026.100860. Contributor Information Xuanhao Wang, Email: xh-wang@zhejianglab.org. Junhui Shi, Email: junhuishi@outlook.com. Supplementary material Supplementary material Data availability Data will be made available on request. References Associated Data Supplementary Materials Supplementary material Data Availability Statement Data will be made available on request.