Pith. sign in

REVIEW 3 major objections 4 minor 35 references

A Transformer that treats each ultrasound sensor's time series as a token reconstructs photoacoustic images directly from raw measurements, bypassing the system matrix at inference and cutting reconstruction time by an order of magnitude.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-01 01:59 UTC pith:HHG6S3RY

load-bearing objection The sensor-token attention idea is clean and the paper is unusually honest about its limits, but every training and test sample comes from one approximate forward model, so the empirical claims are only as strong as that model's r≈0.92 validation. the 3 major comments →

arxiv 2607.25576 v1 pith:HHG6S3RY submitted 2026-07-28 cs.AI

Matrix-Free Photoacoustic Image Reconstruction via Sensor-Token Self-Attention

Mary John , Shibili Said , Imad Barhumi , Sherzod Turaev , Mohamed Yahia This is my paper
classification cs.AI
keywords photoacoustic tomographysparse-view imagingTransformerself-attentionsensor tokenssystem-matrix-free reconstructioncompressive sensingdeep learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper is trying to establish that photoacoustic tomography can be done in real time by replacing iterative inversion with a learned, system-matrix-free mapping. It builds a Transformer whose input tokens are the full time series recorded by each ultrasound sensor; self-attention lets the network exploit cross-sensor acoustic correlations that the forward model encodes, and a learned decoder produces the image in one forward pass. On 46 held-out simulated phantoms, this Sensor Attention Network reports the highest mean SSIM (0.522), the highest mean PSNR (22.09 dB), and the lowest mean NMSE (0.233) among four methods, with the gains over the strongest iterative baseline significant at p<1e-8 on PSNR, NMSE, and Pearson correlation. If the claim holds, real-time clinical PAT becomes feasible because inference no longer multiplies by the large system matrix.

Core claim

The authors construct an analytical k-space forward operator H (an 11,360-by-4,096 matrix for a 64-by-64 grid with 71 sensors and 160 time samples) by factoring it as a time-domain propagation kernel times a 2-D DFT matrix, and regularize it with Hann k-space apodization plus Gaussian temporal damping; together these reduce the energy-normalized mismatch against an independent pseudo-spectral solver from 0.496 to 0.254 (a 49 percent reduction), at a mean per-sensor Pearson correlation of 0.919 ± 0.049. They then train the Sensor Attention Network, a Transformer whose tokens are the full time series of each sensor, followed by a three-layer fully connected spatial decoder, using a vessel-weig

What carries the argument

The central object is the analytical k-space H-matrix, factored as H = K W_fwd, built from two-dimensional discrete Fourier transforms and a cosine propagation kernel, and regularized by Hann apodization and Gaussian temporal damping. The carrying mechanism is the Sensor Attention Network: each sensor's N_t-length time series becomes a token, projected to d_model=128 embeddings plus sinusoidal positional encodings, processed by two stacked multi-head self-attention blocks (four heads) with residual connections and layer norm, and decoded by three fully connected layers with sigmoid activation to a 64-by-64 image. A vessel-weighted MSE loss (weight 8 on vessel pixels, 1 on background) drives

Load-bearing premise

The load-bearing premise is that the analytical k-space forward model used to synthesize all training and test data truly represents photoacoustic physics, even though its agreement with an independent numerical simulation is only about 0.92 per-sensor correlation, not near 1; if that premise fails, every reported comparison inherits the same systematic error.

What would settle it

Acquire experimental photoacoustic signals from a physical phantom with a known optical-absorption map, or synthesize measurements with an independent higher-fidelity solver that includes absorbing boundaries, attenuation, and acoustic heterogeneity, and rerun all four reconstruction methods; if SAN's PSNR and NMSE advantages over LISTA shrink or reverse, the reported result is an artifact of training and testing on the same analytical H-matrix. A simpler check: truncate the final 20 percent of each sensor time series, where the analytical model's mismatch concentrates, and see whether SAN's r

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single GPU forward pass replaces hundreds of matrix-vector products, putting PAT reconstruction within a real-time clinical latency budget (reported 1–5 ms per image).
  • The same validated analytical H-matrix used to generate training data can serve any downstream reconstruction architecture, not just SAN.
  • Because inference no longer scales with unrolled depth times O(MN), the method's cost is decoupled from the system-matrix dimension.
  • The regularized H-matrix compacts the singular-value spectrum by 3.3x, which would also accelerate classical iterative solvers that still use the matrix.
  • The paper argues the sensor-as-token paradigm is not PAT-specific and can extend to other fixed-array tomographic modalities and to 3D volumetric reconstruction.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • An untested consequence: because every method is trained and evaluated on the same synthetic forward model, real experimental data could shrink or reverse the reported margins if the analytical model's ~0.92 per-sensor correlation reflects systematic errors that the network has learned to exploit.
  • The equal SSIM but higher PSNR, NMSE, and Pearson correlation of SAN versus LISTA suggests the network preserves amplitude and full vessel coverage rather than just boundary sharpness; a test on phantoms with vessel topologies far from the training distribution would reveal whether this is learned or general.
  • The paper notes that a physically exact augmentation (rotating or reflecting the phantom while permuting sensor indices accordingly) was not used; applying that augmentation would be a direct test of whether the reported gains survive a realistic distribution shift.
  • A hybrid design that initializes a few unrolled iterations with SAN's fast output could recover fidelity lost to the synthetic forward model while keeping latency near real time; the paper lists physics-informed losses as future work but does not test such a hybrid.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes the Sensor Attention Network (SAN), a Transformer encoder that treats each sensor's full time series as a token and maps raw PAT measurements directly to a reconstructed image, avoiding any H-matrix multiplication at inference. An analytical k-space forward operator H is constructed, regularized with Hann apodization and Gaussian temporal damping, and validated against k-Wave simulations (mean per-sensor Pearson r = 0.919 ± 0.049). The network is trained with a vessel-weighted loss on 488 augmented Radiopaedia-derived phantoms and evaluated on 46 held-out simulated samples against ISTA, SBTV, and LISTA. The paper reports SAN achieves the highest mean SSIM (0.522), PSNR (22.09 dB), and lowest NMSE (0.233), with statistically significant improvements over LISTA on PSNR/NMSE/Pearson at p < 1e-8, while being statistically indistinguishable from LISTA on SSIM. Section 6.3 acknowledges that generalization to experimental measurements is unproven because all data are generated by the analytical H-matrix.

Significance. If the empirical claims hold, the paper makes a useful contribution to fast PAT reconstruction: a direct, non-iterative architecture that removes the system-matrix cost at inference, an analytical forward model with explicit validation against an external solver, and a careful paired statistical comparison with confidence intervals. The transparent validation of H against k-Wave and the honest limitation statement in §6.3 are strengths. However, the significance is currently bounded by the simulation setup: every training and test sample is produced by the same approximate operator that the network learns to invert, and the external validation has moderate correlation (r = 0.919) with a non-negligible NMSE (0.254). The headline superiority over LISTA and classical solvers is therefore established for a self-generated, approximate forward model rather than for physical PAT data.

major comments (3)
  1. [§5.5 and §6.3] All training and test sensor data are generated by the analytical H-matrix of Eq. (3) (§4.7), and the same H is used as the forward operator for ISTA/SBTV/LISTA. The only external check, §5.1, gives mean per-sensor Pearson r = 0.919 ± 0.049 and an energy-normalized NMSE of 0.254 after parameter tuning (§5.2.2). This is an inverse-crime setup: the reported superiority of SAN is demonstrated on data synthesized by the very operator the network was trained to invert, and §6.3 concedes that the gap to experimental measurements is unquantified. A concrete, feasible test is to generate the 46-sample test set with k-Wave using the same geometry and phantoms and re-run all four methods. Without such a test, the abstract claim that SAN achieves the highest fidelity on held-out PAT data is not supported beyond the analytical simulation regime.
  2. [§5.1 vs §5.2.3] The primary validation reports mean per-sensor correlation r = 0.919 ± 0.049 for a centrally placed source, while §5.2.3 reports r = 0.9907 for centre (32,32). If these refer to the same configuration, the numbers are contradictory; if §5.2.3 uses a point source and §5.1 uses a different phantom or source, that distinction must be stated explicitly. This inconsistency matters because the validation correlation is the central evidence that H is a faithful substitute for k-Wave.
  3. [§3.3 and Table 1] Eq. (5) defines the safety factor as ξ = 1.25 and, with T = 6 µs and dt_max = 4.71e-8 s, gives N_t ≈ 160. Table 1 instead lists 'Time samples (CFL, ξ=1.5)' with N_t = 160, which would imply N_t ≈ 191. The parameter value in Table 1 should be corrected to 1.25 or the calculation explained. Relatedly, §5.3.1 says ISTA was run up to 500 iterations while §6.2 states 200; §5.3.2 gives SBTV N_out = 76 while §6.2 says 60 outer plus 5 inner CG iterations. These inconsistencies do not alter the qualitative ranking but must be reconciled.
minor comments (4)
  1. [References] Reference [17] is mis-cited; the Transformer/positional-encoding attribution in §4.2 correctly points to [32] (Vaswani et al.), but §2.4 cites [28] for 'Transformer-based architecture', and [28] is a diffusion-comparison paper. Please correct these citation mappings.
  2. [§3.4, Eqs. (6a)-(6b)] The Hadamard-product notation is unclear: W_fwd is (N_out^2 × N_x N_y) and W_inv is (N_s × N_out^2), while w(k) is a 2D k-space window. State explicitly how the window is broadcast or applied (e.g., row-wise scaling of the DFT matrices) so the construction is reproducible.
  3. [§5.4] The coefficient of determination R^2 = 0.654 ± 0.237 is reported but never defined or included in Table 3. Specify how R^2 is computed (per-sample or pooled) and why it is omitted from the main comparison.
  4. [Grammar/typos] Minor language issues: 'hows' in §5.2.2, 'The test images used is shown' in §§5.3.1–5.3.3, and 'a consequence' hyphenation in §5.4.2. These do not affect the technical content.

Circularity Check

0 steps flagged

No significant circularity: the empirical comparison is a self-contained simulation benchmark using an independently validated forward operator, with explicit limitation statements.

full rationale

The claimed derivation chain is not circular. The H-matrix is constructed analytically (Eqs. 2-3), then validated against an external numerical solver, k-Wave (Sec. 5.1, r=0.919), which is genuine independent evidence. The same H is used to generate training data, test data, and to run iterative baselines; this is a standard simulation benchmark and does not make any fitted parameter equivalent to a predicted quantity by construction. The paper explicitly acknowledges the limitation: 'the gap between the analytical training distribution and experimental measurements must be quantified before clinical translation' (Sec. 6.3). Self-citations to prior work on ISTA, SBTV, and unrolled networks are contextual, not load-bearing; no uniqueness theorem or ansatz is smuggled in via self-citation. The reconstruction superiority claim is conditional on H being a faithful forward model, which is a limitation of external validity, not circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 1 invented entities

The paper's central numbers depend on six hand-selected or fitted parameters (grid spacing, damping time, window choice, two baseline regularisation parameters, loss weights), on the homogeneous-lossless-medium forward model, and on the self-generated training/evaluation data. The forward model is validated externally against k-Wave, which is the strongest part of the paper, but at a modest correlation level and with parameter selection performed on that validation target. Invented entities are limited to the network architecture itself, which is not a new physical postulate.

free parameters (6)
  • Grid spacing d_x = 0.100 mm
    Selected in Section 5.2.4 from {0.075, 0.100, 0.125} to maximise agreement with k-Wave. This choice sets CFL time steps, N_t, and the entire training-data distribution.
  • Gaussian damping time constant tau = 0.9 T
    Chosen by hand in Section 3.4 (Eq. 7) and validated in the ablation of Section 5.2.2; it substantially changes the training-data distribution (NMSE 0.420->0.254 when combined with apodization).
  • Apodization window = Hann
    Selected from Hann/Hamming/Blackman in Section 5.2.1 on the basis of mean Pearson correlation (0.9187 vs 0.9183 vs 0.9197); differences are within noise, so the choice is effectively arbitrary.
  • Regularisation parameter lambda (ISTA) = 0.009
    Fixed in Section 5.3.1 without a reported grid search; directly controls ISTA quality.
  • SBTV parameters (lambda, mu, outer iterations) = lambda=0.1, mu=1.0, N_out=76
    Reported as 'optimal parameters identified during the parameter search' in Section 5.3.2, but no search procedure or selection criterion is given.
  • Loss weights (vessel vs background) = 8 and 1, threshold 0.1
    Hand-chosen in Section 4.6 (Eq. 16); no ablation.
axioms (4)
  • domain assumption Photoacoustic forward problem is exactly y = H x0 + n under a homogeneous, lossless medium with constant sound speed.
    Invoked in Eq. (1) and in the H-matrix construction (Section 3.2); the basis for the whole simulated-data design. Explicitly acknowledged as a limitation in Section 6.3.
  • domain assumption The analytical k-space H-matrix (with apodization and Gaussian damping) is a faithful surrogate for k-Wave simulations for training and evaluation.
    Claimed on the basis of mean Pearson correlation r=0.919 (Section 5.1). The residual mismatch is not quantified in terms of how it propagates into reconstruction metrics.
  • domain assumption Radiopaedia images are representative of PAT-relevant vessel structures, and translation-only augmentation (offsets of 2-3 pixels, horizontal/vertical/diagonal) spans the geometry distribution relevant to the task.
    Dataset construction in Section 4.7 and Discussion 6.3; the paper acknowledges generalisation to other topologies and scales is unproven.
  • domain assumption Pearson correlation is an adequate measure of forward-model fidelity.
    Primary figure of merit in Eq. (8) and throughout Section 5.2. Pearson ignores amplitude and energy errors; the paper separately reports NMSE but uses only correlation for parameter selection.
invented entities (1)
  • Sensor Attention Network (SAN) architecture no independent evidence
    purpose: Maps raw sensor time series to reconstructed PAT image in a single forward pass without the H-matrix.
    A new neural architecture, not a new physical entity. It has no falsifiable handle outside the paper's own experiments.

pith-pipeline@v1.3.0-alltime-deepseek · 17675 in / 9539 out tokens · 67769 ms · 2026-08-01T01:59:02.451621+00:00 · methodology

0 comments
read the original abstract

Photoacoustic tomography (PAT) combines the optical absorption contrast of biological tissue with the spatial resolution of ultrasound, yet recovering the initial pressure distribution from sparse-view sensor measurements remains an ill-posed inverse problem. Iterative compressive-sensing solvers and unrolled deep networks both retain a dependence on the system matrix at inference, which leaves real-time clinical reconstruction computationally expensive. This paper proposes the Sensor Attention Network (SAN), a Transformer-based architecture that treats the full time series of each sensor as a token and maps raw measurements directly to the reconstructed image without invoking the system matrix at inference. For training and benchmarking, an analytical k-space H-matrix is constructed and validated against the k-Wave pseudo-spectral solver under matched geometry, achieving a mean per-sensor Pearson correlation of 0.919 +/- 0.049, with k-space apodization and Gaussian temporal damping acting synergistically to reduce the energy-normalized mismatch by 49%. Trained with a vessel-weighted loss on 488 augmented samples and evaluated on 46 held-out samples against ISTA, split-Bregman total variation (SBTV), and learned ISTA (LISTA), SAN attains the highest mean SSIM (0.522) and PSNR (22.09 dB) and the lowest NMSE (0.233). Paired t-tests and Wilcoxon signed-rank tests confirm the superiority of SAN over LISTA on PSNR, NMSE, and Pearson correlation at p < 1e-8, and over ISTA and SBTV on all fidelity metrics. By bypassing the H-matrix at inference, SAN reduces reconstruction time by at least an order of magnitude, supporting real-time PAT reconstruction.

Figures

Figures reproduced from arXiv: 2607.25576 by Imad Barhumi, Mary John, Mohamed Yahia, Sherzod Turaev, Shibili Said.

Figure 1
Figure 1. Figure 1: PAT System Geometry: imaging grid (64×64) and square sensor array (Ns = 71). dx = 0.100mm. To ensure that all wavefronts originating within the imaging domain reach the sensor bound￾ary before the simulation terminates, the total measurement duration T and the number of time samples Nt are determined jointly as Nt = T ξ dt,max , (5) where ξ = 1.25 is a safety factor that provides a temporal margin to preve… view at source ↗
Figure 2
Figure 2. Figure 2: SAN Pipeline 10 [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Waveform overlays for five randomly selected sensors and their corresponding corre [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Correlation histogram and distribution. rs = 0.80. The full per-sensor correlation profile is shown in [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Mean Pearson correlation achieved by each windowing techniques and the correspond [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Mean correlation for a point source in the centre, off-centre and near-edge (left); [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The metric distribution histograms for all three methods (y-axis is count). [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Some examples of test images used [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: ISTA reconstruction results. 5.3.2 SBTV SBTV was run with the optimal parameters identified during the parameter search (λ = 0.1000, µ = 1.0, Nout = 76 outer iterations). Across the 46 test samples, SBTV achieved a mean SSIM of 0.1733 ± 0.0750, PSNR of 9.36 ± 1.65 dB, NMSE of 4.1700 ± 6.0986, and r¯img of 0.8557±0.0686. SBTV performed below ISTA on all four metrics. The lower Pearson correlation relative t… view at source ↗
Figure 10
Figure 10. Figure 10: SBTV reconstruction results 16 [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: LISTA reconstruction results [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Per-sample SSIM distribution of the SAN reconstructions on the 46-sample test set. [PITH_FULL_IMAGE:figures/full_fig_p018_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Six random test-set reconstructions well-represented in the training distribution and less well on outlier topologies, as visible in the lower tail of the distribution in [PITH_FULL_IMAGE:figures/full_fig_p018_13.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 10 canonical work pages · 2 internal anchors

  1. [1]

    Beard, Biomedical photoacoustic imaging, Interface focus 1 (4) (2011) 602–631.doi: 10.1098/rsfs.2011.0028

    P. Beard, Biomedical photoacoustic imaging, Interface focus 1 (4) (2011) 602–631.doi: 10.1098/rsfs.2011.0028

    arXiv 2011

  2. [2]

    L. Li, L. V. Wang, Recent advances in photoacoustic tomography, BME Frontiers 2021 (2021) 9823268.doi:10.34133/2021/9823268

    work page doi:10.34133/2021/9823268 2021

  3. [3]

    Zhang, F

    J. Zhang, F. Duan, Y. Liu, L. Nie, High-resolution photoacoustic tomography for early- stage cancer detection and its clinical translation, Radiology: Imaging Cancer 2 (3) (2020) e190030, pMID: 33778711.doi:10.1148/rycan.2020190030

    work page doi:10.1148/rycan.2020190030 2020

  4. [4]

    G. P. Luke, D. Yeager, S. Y. Emelianov, Biomedical applications of photoacoustic imaging with exogenous contrast agents, Annals of biomedical engineering 40 (2) (2012) 422–437. doi:10.1007/s10439-011-0449-4

    work page doi:10.1007/s10439-011-0449-4 2012

  5. [5]

    M. John, I. Barhumi, Handling missing data in limited-view photoacoustic tomography using compressive sensing algorithm-based deep learning, in: 2024 Asia Pacific Signal and InformationProcessingAssociationAnnualSummitandConference(APSIPAASC),IEEE, 2024, pp. 1–6.doi:https://doi.org/10.1109/APSIPAASC63619.2025.10848678

    arXiv 2024

  6. [6]

    Model-corrected learned primal-dual models for fast limited-view photoacoustic tomography

    A. Hauptmann, J. Poimala, Model-corrected learned primal-dual models for fast limited- view photoacoustic tomography, arXiv preprint arXiv:2304.01963 (2023).doi:https:// doi.org/10.48550/arXiv.2304.01963

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2304.01963 2023

  7. [7]

    M. J. John, I. Barhumi, Fast and efficient pat image reconstruction algorithms: A comparative performance analysis, Signal Processing 201 (2022) 108691.doi:https: //doi.org/10.1016/j.sigpro.2022.108691

    arXiv 2022

  8. [8]

    M. J. John, I. Barhumi, Total variation algorithms for pat image reconstruction, in: 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Con- ference (APSIPA ASC), IEEE, 2022, pp. 1164–1168.doi:https://doi.org/10.23919/ APSIPAASC55919.2022.9980048

    arXiv 2022

  9. [9]

    M. J. John, I. Barhumi, Compressive sensing based algorithms for limited-view pat image reconstruction, in: 2023 Asia pacific signal and information processing association annual summit and conference (APSIPA ASC), IEEE, 2023, pp. 1317–1322.doi:https://doi. org/10.1109/APSIPAASC58517.2023.10317590

    arXiv 2023

  10. [10]

    M. John, I. Barhumi, Plug-and-play enhanced compressive sensing for limited sample pat image reconstruction, in: 2023 6th international conference on signal processing and in- formation security (ICSPIS), IEEE, 2023, pp. 110–115.doi:https://doi.org/10.1109/ ICSPIS60075.2023.10343788

    arXiv 2023

  11. [11]

    M. John, I. Barhumi, Advancing sensor-data based pat image reconstruction through effi- cient and intelligible unrolled networks, Ieee Access 11 (2023) 117053–117066.doi:https: //doi.org/10.1109/ACCESS.2023.3326504

    arXiv 2023

  12. [12]

    Shahid, A

    H. Shahid, A. Khalid, X. Liu, M. Irfan, D. Ta, A deep learning approach for the photoa- coustic tomography recovery from undersampled measurements, Frontiers in Neuroscience 15 (2021) 598693.doi:https://doi.org/10.3389/fnins.2021.598693

    arXiv 2021

  13. [13]

    Antholzer, M

    S. Antholzer, M. Haltmeier, J. Schwab, Deep learning for photoacoustic tomography from sparse data, Inverse problems in science and engineering 27 (7) (2019) 987–1005.doi: https://doi.org/10.1080/17415977.2018.1518444. 23

    arXiv 2019

  14. [14]

    M. John, I. Barhumi, Unrolled deep learning for breast cancer detection using limited-view photoacoustic tomography data, Medical & Biological Engineering & Computing 63 (6) (2025) 1777–1795.doi:https://doi.org/10.1007/s11517-025-03302-4

    work page doi:10.1007/s11517-025-03302-4 2025

  15. [15]

    Y. Li, O. Bar-Shira, V. Monga, Y. C. Eldar, Deep algorithm unrolling for biomedical imag- ing, arXiv preprint arXiv:2108.06637 (2021).doi:https://doi.org/10.48550/arXiv. 2108.06637

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2108.06637 2021

  16. [16]

    M. A. Josy, Optimizing photoacoustic imaging with advanced reconstruction techniques: Compressive sensing and deep learning, Dissertations (2024). URLhttps://scholarworks.uaeu.ac.ae/all_dissertations/309

    2024

  17. [17]

    Vaswani, Ls-cs-residual (ls-cs): Compressive sensing on least squares residual, IEEE Transactions on Signal Processing 58 (8) (2010) 4108–4120.doi:https://doi.org/10

    N. Vaswani, Ls-cs-residual (ls-cs): Compressive sensing on least squares residual, IEEE Transactions on Signal Processing 58 (8) (2010) 4108–4120.doi:https://doi.org/10. 1109/TSP.2010.2048105

    arXiv 2010

  18. [18]

    M. Xu, L. V. Wang, Universal back-projection algorithm for photoacoustic computed tomography, Physical Review E—Statistical, Nonlinear, and Soft Matter Physics 71 (1) (2005) 016706.doi:https://doi.org/10.1103/PhysRevE.71.016706

    work page doi:10.1103/physreve.71.016706 2005

  19. [19]

    A. Beck, M. Teboulle, A fast iterative shrinkage-thresholding algorithm for linear inverse problems, SIAM Journal on Imaging Sciences 2 (1) (2009) 183–202.doi:https://doi. org/10.1137/080716542

    doi:10.1137/080716542 2009

  20. [20]

    Z. Qin, D. Goldfarb, S. Ma, An alternating direction method for total variation denoising, Optimization Methods and Software 30 (3) (2015) 594–615.doi:https://doi.org/10. 1080/10556788.2014.955100

    arXiv 2015

  21. [21]

    S. V. Venkatakrishnan, C. A. Bouman, B. Wohlberg, Plug-and-play priors for model based reconstruction, in: 2013 IEEE global conference on signal and information processing, IEEE, 2013, pp. 945–948.doi:https://doi.org/10.1109/GlobalSIP.2013.6737048

    arXiv 2013

  22. [22]

    U. S. Kamilov, H. Mansour, B. Wohlberg, A plug-and-play priors approach for solving nonlinear imaging inverse problems, IEEE Signal Processing Letters 24 (12) (2017) 1872– 1876.doi:https://doi.org/10.1109/LSP.2017.2763583

    arXiv 2017

  23. [23]

    Awasthi, S

    N. Awasthi, S. Kumar Kalva, M. Pramanik, P. K. Yalavarthy, Dimensionality reduced plug and play priors for improving photoacoustic tomographic imaging with limited noisy data, Biomedical Optics Express 12 (3) (2021) 1320–1338.doi:https://doi.org/10.1364/ BOE.415182

    2021

  24. [24]

    Gregor, Y

    K. Gregor, Y. LeCun, Learning fast approximations of sparse coding, in: Proceedings of the 27th international conference on international conference on machine learning, 2010, pp. 399–406.doi:https://dl.acm.org/doi/abs/10.5555/3104322.3104374

    arXiv 2010

  25. [25]

    H. Lan, D. Jiang, C. Yang, F. Gao, F. Gao, Y-net: Hybrid deep learning image re- construction for photoacoustic tomography in vivo, Photoacoustics 20 (2020) 100197. doi:https://doi.org/10.1016/j.pacs.2020.100197

    arXiv 2020

  26. [26]

    K.-T. Hsu, S. Guan, P. V. Chitnis, Comparing deep learning frameworks for photoacoustic tomography image reconstruction, Photoacoustics 23 (2021) 100271.doi:https://doi. org/10.1016/j.pacs.2021.100271

    arXiv 2021

  27. [27]

    Rajendran, A

    P. Rajendran, A. Sharma, M. Pramanik, Photoacoustic imaging aided with deep learning: a review, Biomedical Engineering Letters 12 (2) (2022) 155–173.doi:https://doi.org/ 10.1007/s13534-021-00210-y. 24

    work page doi:10.1007/s13534-021-00210-y 2022

  28. [28]

    S. Said, I. Barhumi, A comparative study of diffusion-based reconstruction frameworks for photoacoustic tomography, Scientific Reports (2026)

    2026

  29. [29]

    J. Yin, J. He, X. Ma, M. Xie, Optimization research on sparse sampling photoacous- tic tomography based on attention-unet, in: Second International Conference on Optical Communication and Optoelectronic Technology (OCOT 2025), Vol. 13985, SPIE, 2025, pp. 486–493.doi:https://doi.org/10.1117/12.3077798

    work page doi:10.1117/12.3077798 2025

  30. [30]

    Y. Lai, Q. Zhang, Z. Yin, Attention-driven complementary information fusion network for sparse photoacoustic image reconstruction, Photoacoustics (2026) 100797doi:https: //doi.org/10.1016/j.pacs.2026.100797

    arXiv 2026

  31. [31]

    B. E. Treeby, B. T. Cox, k-wave: Matlab toolbox for the simulation and reconstruction of photoacoustic wave fields, Journal of biomedical optics 15 (2) (2010) 021314–021314. doi:https://doi.org/10.1117/1.3360308

    doi:10.1117/1.3360308 2010

  32. [32]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin, Attention is all you need, Advances in neural information processing systems 30 (2017).doi:https://doi.org/10.48550/arXiv.1706.03762

    Pith/arXiv arXiv doi:10.48550/arxiv.1706.03762 2017

  33. [33]

    11, 2026

    Radiopaedia.org, Radiopaedia.org,https://radiopaedia.org, accessed: Jul. 11, 2026

    2026

  34. [34]

    K. He, X. Zhang, S. Ren, J. Sun, Delving deep into rectifiers: Surpassing human-level per- formance on imagenet classification, in: Proceedings of the IEEE international conference on computer vision, 2015, pp. 1026–1034

    2015

  35. [35]

    Huynh, O

    N. Huynh, O. Francies, E. Zhang, J. Zhu, J. Manson, P. Beard, A. A. Plumb, Feasibility of 3d photoacoustic tomography for the assessment of rheumatoid arthritis, British Journal of Radiology 99 (1179) (2026) 473–480.doi:https://doi.org/10.1093/bjr/tqaf313. 25

    work page doi:10.1093/bjr/tqaf313 2026

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.