A reconstruction method called PAGS posted the best reported scores against the study's simulated reference and sharper absorber boundaries in a physical phantom than the listed comparison methods, according to arXiv:2608.25472v1, a preprint dated 26 Aug 2026. In a sparse-view test, major vascular trunks and branch connectivity remained visible when detector retention was 12.5%. The work's validation was limited to simulated and phantom data.
The method targets PACT reconstruction under heterogeneous speed-of-sound conditions, where the sound speed varies across the medium. PAGS is designed for blind autofocusing, meaning it learns a path-averaged speed-of-sound correction instead of relying on calibrated speed-of-sound priors. The authors describe that correction as an effective path-level propagation field rather than a recovered physical local speed-of-sound map.
Learning the missing sound-speed correction
At the model's core are Gaussian PA sources and a direction-varying ASoS field, the study's learned path-level sound-speed representation. It interpolates first across space and then across direction with SH coefficients, while an analytic acoustic forward model synthesizes transducer measurements. Joint optimization runs through the full computation graph: Adam updates the source parameters and ASoS SH coefficients, while density control splits high-gradient sources and prunes negligible ones.
The in-silico test used k-Wave-simulated measurements from a 3D vascular phantom with a dual-region medium. The background speed of sound was 1,450 m/s, while an ellipsoidal inclusion was set at 1,550 m/s and had semi-axes of 7, 8 and 11 mm. The acquisition used 4,600 point transducers, 1,024 samples per detector and a 50 MHz sampling rate.
The physical phantom contained absorptive structures and a tissue region with approximately 5% to 8% acoustic contrast to water. Its measurements used 575 elements over eight rotational views, yielding 4,600 effective positions, with 4,096 samples per channel at 20 MHz.
What the phantom tests showed
For the simulation, reconstructions were peak-normalized and rigidly registered, with the transform estimated by maximizing normalized cross-correlation. Scores used PSNR, RMSE and SSIM against a constructed uniform-SoS UBP reference. Physical-phantom quality was assessed through visual inspection, local contrast and line-profile widths.
Against that reference, PAGS had the best reported combination: PSNR 30.3, RMSE 0.0305 and SSIM 0.537. Vanilla SlingBAG scored 29.0, 0.0354 and 0.331, respectively. The reported differences were 1.3 dB in PSNR and 0.206 in SSIM, while RMSE moved from 0.0354 to 0.0305.
In the physical phantom, PAGS showed sharper absorber boundaries, improved local contrast and clearer separation between neighboring structures than the three baselines. The sparse-view test retained 50%, 25% and 12.5% of detectors. At the lowest retention level, major vascular trunks and branch connectivity remained visible, with degradation appearing mainly as reduced fine-detail contrast rather than severe streak artifacts.
Efficiency and limits
An ablation study compared signal-domain residual convergence across matched configurations. The ASoS-containing configurations had lower final residuals; the analytic and 10-sphere models showed similar convergence trends, and full PAGS reached the lowest final residual. Numerical final residual values and uncertainty estimates were not reported.
Runtime comparisons were made under identical hardware and problem scale. Reported per-iteration times were 6.7 seconds for the baseline, 6.9 seconds for 10-sphere plus SH probes, 2.3 seconds for analytic projection with uniform SoS and 2.4 seconds for full PAGS. The analytic configurations were faster per iteration, and adding ASoS changed the analytic setting only from 2.3 to 2.4 seconds.
The paper also reports a memory-efficiency calculation for storing and querying the ASoS field. In the representative calculation, the probe grid required 36.9K learnable weights, a reported 12.5K-fold reduction in parameters. Spatial-first querying reduced the intermediate cache from 460M to 900K values, about 510-fold. The paper presents these as representative figures, not a general empirical guarantee.
The main caution is the scope of the validation. PAGS estimates an effective path-level propagation field rather than recovering a physical SoS map, and the reported validation is limited to simulated and phantom data. The authors identify broader testing in more complex acoustic settings as future work.
The authors say the analyzed datasets are available from the corresponding author on reasonable request. The official implementation, usage instructions and a synthetic example are publicly available at https://github.com/work-submit/PAGS/.
Paper data and sources
Original title: PAGS: Autofocusing Photoacoustic Tomography via Speed-of-Sound-Adaptive Gaussian Splatting
Authors: Jiarui Ge, Jintao Ma, Bangxu Fan et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text