pith. sign in

arxiv: 2604.19176 · v1 · submitted 2026-04-21 · 📡 eess.IV · cs.LG· math.OC

Deep Image Prior for photoacoustic tomography can mitigate limited-view artifacts

Hanna Pulkkinen , Jenni Poimala , Leonid Kunyansky , Janek Gr\"ohl , Andreas Hauptmann This is my paper

Pith reviewed 2026-05-10 01:56 UTC · model grok-4.3

classification 📡 eess.IV cs.LGmath.OC
keywords photoacoustic tomographydeep image priorlimited-view artifactsunsupervised reconstructiontotal variation regularizationimage reconstructioncircular measurement geometry
0
0 comments X

The pith

Deep image prior reconstruction mitigates limited-view artifacts in photoacoustic tomography better than total variation methods.

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

The paper tests whether the deep image prior can serve as an unsupervised reconstruction method for photoacoustic tomography. It focuses on the common experimental problem of limited viewing angles that produce artifacts and noise in the resulting images. By pairing the prior with fast forward and adjoint operators for circular geometries, a fast initial inverse, and total variation regularization to control overfitting, the approach is compared directly against classical total variation reconstruction. Experiments on simulated data with added noise and on real measurements validated by a digital twin show quantitative gains in several image quality metrics. A sympathetic reader would care because photoacoustic tomography is used for medical and biological imaging where full-view access is often impractical, so an unsupervised fix could improve reliability without large training datasets.

Core claim

The deep image prior framework, when initialized via a fast inverse and combined with total variation regularization, provides an effective unsupervised strategy for robust photoacoustic tomography reconstruction even under limited-view geometries. Using recently published fast forward and adjoint algorithms for circular measurement setups, the method reduces artifacts and noise on both simulated measurements with varying noise levels and experimental data assessed via a digital twin, yielding improvements over classical total variation reconstructions in quantitative measures.

What carries the argument

The deep image prior optimization applied to the photoacoustic inverse problem, driven by fast circular-geometry operators and stabilized by total variation regularization plus fast-inverse initialization.

If this is right

  • Unsupervised reconstruction becomes feasible for limited-view photoacoustic tomography without requiring paired training data.
  • Image quality improves in both simulated noisy conditions and real experimental setups relative to total variation baselines.
  • Fast circular-geometry operators make the method computationally practical for standard measurement configurations.
  • The combination of initialization and regularization extends the usable range of limited-view geometries in practice.

Where Pith is reading between the lines

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

  • Similar unsupervised priors could be tested on other tomographic modalities that suffer from restricted angular coverage.
  • Adjusting the regularization strength or initialization method might further improve robustness under higher noise levels.
  • Adoption could lower hardware requirements for full angular sampling in photoacoustic systems.

Load-bearing premise

Initialization with a fast inverse plus total variation regularization prevents the deep image prior optimization from overfitting to noise or limited-view artifacts on photoacoustic data.

What would settle it

Quantitative comparison of deep image prior and total variation reconstructions on additional experimental limited-view photoacoustic data with an independent ground-truth phantom that reveals whether the reported metric gains persist or new artifacts emerge.

Figures

Figures reproduced from arXiv: 2604.19176 by Andreas Hauptmann, Hanna Pulkkinen, Janek Gr\"ohl, Jenni Poimala, Leonid Kunyansky.

Figure 1
Figure 1. Figure 1: Schematic visualization of the MSOT inVision 256-TF by iThera Medical. The data are acquired by a ring array of 256 transducer elements that have an angular coverage of 270◦ . The imaging target is positioned in the center of the ring array and is illuminated from five directions in 72◦ intervals, leading to a homogeneous radiant exposure in the imaging plane. To reflect real-world limitations in sensor ge… view at source ↗
Figure 2
Figure 2. Figure 2: Method pipeline for experimental data. Time-series data is measured using a circular detection geometry and mapped to an initial pressure estimate using a fast inverse operator. This initial reconstruction is refined using the DIP framework, where an untrained U-Net is optimized by minimizing a data-fidelity term together with a total-variation regularization term. The final re￾construction fb is obtained … view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of reconstruction quality using DIP and TV regularization under 270° detector coverage at varying levels of added relative L2-noise (0%, 10%, and 20%). Each row corre￾sponds to a different noise level, while columns represent: (from left to right) initial reconstruction, DIP-based reconstruction with early stopping and running to convergence, and TV-regularized re￾construction. Both methods shar… view at source ↗
Figure 4
Figure 4. Figure 4: PSNR during reconstructions as a function of iteration of DIP and TV method under varying noise: (a) noise-free case, (b) 10% added noise, and (c) 20% added noise. Vertical lines indicate the early-stopping point for DIP and the convergence points for both of the methods [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: SSIM during DIP training under different noise levels (0%, 10% and 20%). The curves show rapid improvement in the early iterations, followed by a gradual stabilization. Higher noise levels result in earlier degradation of performance and lower peak SSIM value [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of reconstruction quality with all 256 de￾tectors in use (270◦ angular coverage). 3.2.1. 256/256 detectors in use (270◦ angular coverage). The results for the full detector geometry (270◦) are shown in [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of reconstruction quality with 170 detec￾tors out of the total 256 in use (179.3◦ angular coverage). 3.2.2. 170/256 detectors in use (179.3 ◦ angular coverage). The results for the 179.3 ◦ case are shown in [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of reconstruction quality with 112 detec￾tors out of the total 256 in use (118.125◦ angular coverage). 3.2.3. 112/256 detectors in use (118.125◦ angular coverage). Finally, we show the results for the strongest limited view in [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Mouse data with 270◦ detector coverage comparing ini￾tial reconstruction with two DIP reconstructions: one with leaky ReLU activation function and one without any nonlinearity in the output layer. DIP used 600 iterations and λ = 0.001 in all cases. TV uses α = 2.0 for all cases, and 2964, 3080, 2964 iterations for Scan 1, Scan 2 and Scan 3, respectively. 4. Discussion. In this study, we compared two recons… view at source ↗
read the original abstract

We study the deep image prior (DIP) framework applied to photoacoustic tomography (PAT) as an unsupervised reconstruction approach to mitigate limited-view artifacts and noise commonly encountered in experimental settings. Efficient implementation is achieved by employing recently published fast forward and adjoint algorithms for circular measurement geometries. Initialization via a fast inverse and total variation (TV) regularization are applied to further suppress noise and mitigate overfitting. For comparison, we compute a classical TV reconstruction. Our experiments comprise simulated PAT measurements under limited-view geometries and varying levels of added noise as well as experimental measurements together with using a digital twin for quality assessment. Our findings suggest that DIP framework provides an effective unsupervised strategy for robust PAT reconstruction even in the challenging case of a limited view geometry providing improvement in several quantitative measures over total variation reconstructions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript applies the Deep Image Prior (DIP) as an unsupervised reconstruction method for photoacoustic tomography (PAT) under limited-view geometries. It leverages recently published fast forward and adjoint operators for circular measurement setups, initializes the network with a fast inverse solution, and adds total variation (TV) regularization during optimization to control noise and overfitting. The approach is tested on simulated limited-view PAT data with varying noise levels and on experimental measurements, with a digital twin used for quantitative quality assessment. The central claim is that this DIP-based strategy yields improved quantitative metrics relative to classical TV reconstruction.

Significance. If the results are robust, the work demonstrates a practical unsupervised route to artifact reduction in PAT that avoids the need for large supervised training sets. The emphasis on efficient forward/adjoint operators for circular geometries and the use of a digital twin for validation are concrete strengths that support reproducibility and applicability in experimental biomedical imaging.

major comments (1)
  1. [Experiments and Results] The experimental comparisons (described in the abstract and results) report gains over TV reconstruction but contain no ablation removing the TV term, altering the fast-inverse initialization, or replacing DIP with a plain iterative solver under the same regularization. Without these controls it is not possible to attribute the reported metric improvements specifically to the DIP implicit prior mitigating limited-view artifacts rather than to the explicit TV regularization and initialization that are already known to suppress artifacts in classical PAT.
minor comments (2)
  1. [Abstract] The abstract states that 'several quantitative measures' improve but does not name them (e.g., PSNR, SSIM, or structural similarity indices) or report numerical values with error bars; adding this information would strengthen the claim.
  2. [Methods] Notation for the fast forward and adjoint operators is introduced without an explicit equation reference or complexity statement, making it harder for readers to verify the claimed efficiency gain.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive review and the opportunity to clarify the experimental design. We address the major comment on ablation studies below and have incorporated revisions to strengthen the attribution of results.

read point-by-point responses

  1. Referee: [Experiments and Results] The experimental comparisons (described in the abstract and results) report gains over TV reconstruction but contain no ablation removing the TV term, altering the fast-inverse initialization, or replacing DIP with a plain iterative solver under the same regularization. Without these controls it is not possible to attribute the reported metric improvements specifically to the DIP implicit prior mitigating limited-view artifacts rather than to the explicit TV regularization and initialization that are already known to suppress artifacts in classical PAT.

    Authors: We agree that the current comparisons do not fully isolate the contribution of the DIP implicit prior from the TV regularization and fast-inverse initialization. In the revised manuscript we have added a dedicated ablation subsection in the Results. This includes: (i) DIP optimization without the TV term, (ii) DIP with random (instead of fast-inverse) initialization, and (iii) a plain iterative solver using identical TV regularization but without the neural-network parameterization. Quantitative metrics on both the simulated limited-view data (across noise levels) and the experimental measurements (via the digital twin) show that each ablated variant underperforms the full DIP approach, supporting that the network prior supplies additional artifact mitigation. These new figures and tables are now referenced in the abstract and discussion. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the experimental DIP-PAT study

full rationale

The paper applies the deep image prior (DIP) method to photoacoustic tomography (PAT) for limited-view reconstruction, using fast algorithms, initialization from a fast inverse, and TV regularization, then compares results to classical TV reconstructions on simulated and experimental data with a digital twin for assessment. The claims are supported by quantitative improvements in metrics rather than any mathematical derivation or prediction that reduces to the inputs by construction. No self-definitional steps, fitted inputs presented as predictions, or load-bearing self-citations that force the result are present in the described approach.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Based solely on the abstract, no explicit free parameters, axioms, or invented entities are detailed; the approach inherits standard DIP assumptions and TV regularization without introducing new postulated entities.

pith-pipeline@v0.9.0 · 5446 in / 1040 out tokens · 30688 ms · 2026-05-10T01:56:07.747301+00:00 · methodology

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

45 extracted references · 45 canonical work pages

  1. [1]

    Antholzer, M

    S. Antholzer, M. Haltmeier and J. Schwab, Deep learning for photoacoustic tomography from sparse data,Inverse Problems in Science and Engineering,27(2019), no. 7, 987–1005

    work page 2019

  2. [2]

    S. R. Arridge, M. M. Betcke, B. T. Cox, F. Lucka and B. E. Treeby, On the adjoint operator in photoacoustic tomography,Inverse Problems,32(2016), no. 11, 115012

    work page 2016

  3. [3]

    D. O. Baguer, J. Leuschner and M. Schmidt, Computed tomography reconstruction using deep image prior and learned reconstruction methods,Inverse Problems,36(2020), no. 9, 094004

    work page 2020

  4. [4]

    Barbano, J

    R. Barbano, J. Leuschner, M. Schmidt, A. Denker, A. Hauptmann, P. Maass and B. Jin, An educated warm start for deep image prior-based micro CT reconstruction,IEEE Transactions on Computational Imaging,8(2022), 1210–1222

    work page 2022

  5. [5]

    Beard, Biomedical photoacoustic imaging,Interface Focus,1(2011), no

    P. Beard, Biomedical photoacoustic imaging,Interface Focus,1(2011), no. 4, 602–631

    work page 2011

  6. [6]

    Breger, A

    A. Breger, A. Biguri, M. S. Landman, I. Selby, N. Amberg, E. Brunner, J. Gr¨ ohl, S. Hatamikia, C. Karner, L. Ning et al., A study of why we need to reassess full reference image quality assessment with medical images,Journal of Imaging Informatics in Medicine, (2025), 1–26

    work page 2025

  7. [7]

    Chambolle and T

    A. Chambolle and T. Pock, A first-order primal-dual algorithm for convex problems with applications to imaging,Journal of Mathematical Imaging and Vision,40(2011), no. 1, 120–145

    work page 2011

  8. [8]

    Cheng, W

    D. Cheng, W. Li and T. Feng, Research on Deep Image Prior-Based Unmixing Method for Photoacoustic Imaging,2024 3rd International Conference on Image Processing and Media Computing (ICIPMC), 2024, 353–359. DIP FOR PAT CAN MITIGATE LIMITED-VIEW ARTIF ACTS 17

    work page 2024

  9. [9]

    W. Choi, D. Oh and C. Kim, Practical photoacoustic tomography: realistic limitations and technical solutions,Journal of Applied Physics,127(2020), no. 23

    work page 2020

  10. [10]

    B. T. Cox and P. C. Beard, Fast calculation of pulsed photoacoustic fields in fluids using k-space methods,The Journal of the Acoustical Society of America,117(2005), no. 6, 3616– 3627

    work page 2005

  11. [11]

    Dittmer, T

    S. Dittmer, T. Kluth, P. Maass and D. O. Baguer, Regularization by architecture: A deep prior approach for inverse problems,Journal of Mathematical Imaging and Vision,62(2020), no. 3, 456–470

    work page 2020

  12. [12]

    Dreier, S

    F. Dreier, S. Pereverzyev Jr and M. Haltmeier, Operator learning approach for the limited view problem in photoacoustic tomography,Computational Methods in Applied Mathematics, 19(2019), no. 4, 749–764

    work page 2019

  13. [13]

    Fang and D

    Q. Fang and D. A. Boas, Monte Carlo simulation of photon migration in 3D turbid media accelerated by graphics processing units,Optics Express,17(2009), no. 22, 20178–20190

    work page 2009

  14. [14]

    Finch and S

    D. Finch and S. K. Patch, Determining a function from its mean values over a family of spheres,SIAM Journal on Mathematical Analysis,35(2004), no. 5, 1213–1240

    work page 2004

  15. [15]

    Gr¨ ohl, M

    J. Gr¨ ohl, M. Schellenberg, K. Dreher and L. Maier-Hein, Deep learning for biomedical pho- toacoustic imaging: A review,Photoacoustics,22(2021), 100241

    work page 2021

  16. [16]

    Gr¨ ohl, K

    J. Gr¨ ohl, K. K. Dreher, M. Schellenberg, T. Rix, N. Holzwarth, P. Vieten, L. Ayala, S. E. Bohndiek, A. Seitel and L. Maier-Hein, SIMPA: an open-source toolkit for simulation and image processing for photonics and acoustics,Journal of Biomedical Optics,27(2022), no. 8, 083010–083010

    work page 2022

  17. [17]

    Gr¨ ohl, T

    J. Gr¨ ohl, T. R. Else, L. Hacker, E. V. Bunce, P. W. Sweeney and S. E. Bohndiek, Moving beyond simulation: data-driven quantitative photoacoustic imaging using tissue-mimicking phantoms,IEEE Transactions on Medical Imaging,43(2023), no. 3, 1214–1224

    work page 2023

  18. [18]

    Gr¨ ohl, K

    J. Gr¨ ohl, K. Yeung, K. Gu, T. R. Else, M. Golinska, E. V. Bunce, L. Hacker and S. E. Bohndiek, Distribution-informed and wavelength-flexible data-driven photoacoustic oximetry, Journal of Biomedical Optics,29(2024), no. S3, S33303–S33303

    work page 2024

  19. [19]

    Gr¨ ohl, L

    J. Gr¨ ohl, L. Kunyansky, J. Poimala, T. R. Else, F. Di Cecio, S. E. Bohndiek, B. T. Cox and A. Hauptmann, Digital twins enable full-reference quality assessment of photoacoustic image reconstructions,Journal of the Acoustical Society of America,158(2025), no. 1, 590–601

    work page 2025

  20. [20]

    S. Guan, A. A. Khan, S. Sikdar and P. V. Chitnis, Limited-view and sparse photoacoustic tomography for neuroimaging with deep learning,Scientific Reports,10(2020), no. 1, 8510

    work page 2020

  21. [21]

    Haltmeier, O

    M. Haltmeier, O. Scherzer and G. Zangerl, A reconstruction algorithm for photoacoustic imaging based on the nonuniform FFT,IEEE Transactions on Medical Imaging,28(2009), no. 11, 1727–1735

    work page 2009

  22. [22]

    Haltmeier and L

    M. Haltmeier and L. V. Nguyen, Analysis of iterative methods in photoacoustic tomography with variable sound speed,SIAM Journal on Imaging Sciences,10(2017), no. 2, 751–781

    work page 2017

  23. [23]

    Hauptmann, F

    A. Hauptmann, F. Lucka, M. Betcke, N. Huynh, J. Adler, B. Cox, P. Beard, S. Ourselin and S. Arridge, Model-based learning for accelerated, limited-view 3-D photoacoustic tomography, IEEE Transactions on Medical Imaging,37(2018), no. 6, 1382–1393

    work page 2018

  24. [24]

    Hauptmann and B

    A. Hauptmann and B. Cox, Deep learning in photoacoustic tomography: current approaches and future directions,Journal of Biomedical Optics,25(2020), no. 11, 112903

    work page 2020

  25. [25]

    Hauptmann and T

    A. Hauptmann and T. Tarvainen, Model-based reconstructions for quantitative imaging in photoacoustic tomography, inBiomedical Photoacoustics: Technology and Applications, Springer, 2024, 133–153

    work page 2024

  26. [26]

    Hauptmann, L

    A. Hauptmann, L. Kunyansky and J. Poimala, Fast algorithms enabling optimization and deep learning for photoacoustic tomography in a circular detection geometry,SIAM Journal on Imaging Sciences, 2026

    work page 2026

  27. [27]

    K.-T. Hsu, S. Guan and P. V. Chitnis, Fast iterative reconstruction for photoacoustic to- mography using learned physical model: Theoretical validation,Photoacoustics,29(2023), 100452

    work page 2023

  28. [28]

    Hwang and G

    G. Hwang and G. Jeon, SelfPAT: Self-supervised joint estimation of sound speed and initial pressure in PAT, Optica Open, 2025

    work page 2025

  29. [29]

    K. H. Jin, M. T. McCann, E. Froustey and M. Unser, Deep convolutional neural network for inverse problems in imaging,IEEE Transactions on Image Processing,26(2017), no. 9, 4509–4522

    work page 2017

  30. [30]

    K. P. K¨ ostli, M. Frenz, H. Bebie and H. P. Weber, Temporal backward projection of optoa- coustic pressure transients using Fourier transform methods,Physics in Medicine & Biology, 46(2001), no. 7, 1863–1872. 18 H. PULKKINEN, J. POIMALA, L. KUNYANKSY, J. GR ¨OHL AND A. HAUPTMANN

    work page 2001

  31. [31]

    L. A. Kunyansky, Explicit inversion formulae for the spherical mean Radon transform,Inverse Problems,23(2007), no. 1, 373–383

    work page 2007

  32. [32]

    Kunyansky, Reconstruction of a function from its spherical (circular) means with the centers lying on the surface of certain polygons and polyhedra,Inverse Problems,27(2011), no

    L. Kunyansky, Reconstruction of a function from its spherical (circular) means with the centers lying on the surface of certain polygons and polyhedra,Inverse Problems,27(2011), no. 2, 025012

    work page 2011

  33. [33]

    Kunyansky, Fast reconstruction algorithms for the thermoacoustic tomography in certain domains with cylindrical or spherical symmetries,Inverse Problems and Imaging,6(2012), no

    L. Kunyansky, Fast reconstruction algorithms for the thermoacoustic tomography in certain domains with cylindrical or spherical symmetries,Inverse Problems and Imaging,6(2012), no. 1, 111–131

    work page 2012

  34. [34]

    H. Lan, J. Zhang, C. Yang and F. Gao, Compressed sensing for photoacoustic computed tomography based on an untrained neural network with a shape prior,Biomedical Optics Express,12(2021), no. 12, 7835–7848

    work page 2021

  35. [35]

    Poudel, Y

    J. Poudel, Y. Lou and M. A. Anastasio, A survey of computational frameworks for solving the acoustic inverse problem in three-dimensional photoacoustic computed tomography,Physics in Medicine & Biology,64(2019), no. 14, 14TR01

    work page 2019

  36. [36]

    M. T. Rietberg, J. Gr¨ ohl, T. R. Else, S. E. Bohndiek, S. Manohar and B. T. Cox, Artifacts in photoacoustic imaging: Origins and mitigations,Photoacoustics, (2025), 100745

    work page 2025

  37. [37]

    Ronneberger, P

    O. Ronneberger, P. Fischer and T. Brox, U-net: Convolutional networks for biomedical im- age segmentation,International Conference on Medical Image Computing and Computer- Assisted Intervention, Springer, 2015, 234–241

    work page 2015

  38. [38]

    F. Rul, M. M. Betcke et al., Hamilton-Green solver for the forward and adjoint problems in photoacoustic tomography,Journal of Computational Physics,449(2022), 110797

    work page 2022

  39. [39]

    K. Tang, S. Zhang, Y. Wang, X. Zhang, Z. Liu, Z. Liang, H. Wang, L. Chen, W. Chen and L. Qi, Learning spatially variant degradation for unsupervised blind photoacoustic tomography image restoration,Photoacoustics,32(2023), 100536

    work page 2023

  40. [40]

    B. E. Treeby and B. T. Cox, k-Wave: MATLAB toolbox for the simulation and reconstruction of photoacoustic wave fields,Journal of Biomedical Optics,15(2010), no. 2, 021314–021314

    work page 2010

  41. [41]

    Ulyanov, A

    D. Ulyanov, A. Vedaldi and V. Lempitsky, Deep image prior,Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, 2018, 9446–9454

    work page 2018

  42. [42]

    T. Vu, A. DiSpirito III, D. Li, Z. Wang, X. Zhu, M. Chen, L. Jiang, D. Zhang, J. Luo, Y. S. Zhang et al., Deep image prior for undersampling high-speed photoacoustic microscopy, Photoacoustics,22(2021), 100266

    work page 2021

  43. [43]

    L. V. Wang et al., Recent advances in photoacoustic tomography,BME Frontiers, 2021

    work page 2021

  44. [44]

    Xu and L

    M. Xu and L. V. Wang, Universal back-projection algorithm for photoacoustic computed tomography,Physical Review E,71(2005), no. 1, 016706

    work page 2005

  45. [45]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman and O. Wang, The unreasonable effectiveness of deep features as a perceptual metric,Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, 586–595. Received xxxx 20xx; revised xxxx 20xx; early access xxxx 20xx

    work page 2018