Pith. sign in

REVIEW 3 major objections 6 minor 55 references

A matrix-free randomize-then-optimize sampler generates independent posterior samples for three-dimensional photoacoustic tomography using only forward and adjoint wave propagations, with an adjoint discretization that is the exact transpos

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 · deepseek-v4-flash

2026-08-01 07:37 UTC pith:HOUMZKFP

load-bearing objection A solid matrix-free RTO/adjoint construction for 3D PAT, but the reported uncertainties rest on an unresolved LSQR stopping issue the authors themselves flag; referee it, revise it. the 3 major comments →

arxiv 2607.21373 v1 pith:HOUMZKFP submitted 2026-07-23 math.NA cs.NAmath.STstat.TH

3D Uncertainty Quantification for the Photo-Acoustic Tomography

Babak Maboudi Afkham , Amal Mohammed A Alghami , Hassan Yazdanian , Tanja Tarvainen This is my paper
classification math.NA cs.NAmath.STstat.TH MSC 68Q2568R1068U05
keywords photoacoustic tomographyBayesian inverse problemsuncertainty quantificationrandomize-then-optimizematrix-free methodsfinite element discretizationadjoint consistencyLSQR
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.

Photoacoustic tomography reconstructs an initial pressure field from ultrasound measurements, and the authors want a trustworthy picture of the uncertainty in that reconstruction for realistic three-dimensional problems. They claim that a matrix-free reformulation of the randomize-then-optimize (RTO) sampler lets them draw independent samples from the Gaussian posterior by solving randomized least-squares problems, each requiring only one forward and one adjoint wave propagation—no assembled system matrices, no direct solves. The central technical move is an adjoint discretization that is the exact algebraic transpose of the discrete forward operator while still converging to the continuous photoacoustic adjoint as the mesh refines, which is what makes LSQR efficient. If the claim holds, Bayesian UQ moves from small two-dimensional test problems to a three-dimensional finite-element geometry with roughly 200,000 unknowns, and the same machinery applies to any linear PDE-constrained inverse problem. The paper also argues that discretization choices—time stepping, absorbing boundary conditions, mass lumping, and prior modeling—materially change the reported posterior variance and must be treated as part of the uncertainty model.

Core claim

For the linear-Gaussian photoacoustic model, the solution of each randomly perturbed weighted least-squares problem is an exact draw from the Gaussian posterior; the mean and covariance identities are shown in Appendix A. The implementation keeps the entire computation matrix-free: the forward map is a mass-lumped Störmer–Verlet time stepping, the adjoint is evaluated by the exact transpose of that discrete propagator via a backward recursion, and LSQR is used to solve each perturbed least-squares problem. The authors verify the transpose identity to machine precision, compare sample statistics against an explicitly assembled posterior covariance on a small mesh, and demonstrate a three-dime

What carries the argument

The central object is the discrete forward operator G formed from repeated application of the one-step Störmer–Verlet propagator with mass lumping, together with its matrix-free transpose. The key identity is that the backward recursion with transposed substeps provides the exact algebraic adjoint—confirmed to machine precision—while remaining consistent with the continuous PAT adjoint once the computational domain is enlarged so boundary reflections cannot re-enter during the observation interval. This exact transpose pair is what lets the randomize-then-optimize sampler solve each perturbed least-squares problem with LSQR rather than a less efficient general solver.

Load-bearing premise

The posterior samples are treated as exact draws from the Gaussian posterior even though each least-squares problem is solved by LSQR with a fixed iteration limit (300 in 2D, 350 in 3D) rather than by a convergence criterion that accounts for both prior and likelihood; if those iterates are not converged, the reported variance maps are biased.

What would settle it

Compare the sample covariance from RTO against the explicitly assembled posterior covariance C_post = (σ^{-2}G^T G + C^{-1})^{-1} on a small mesh, varying the LSQR iteration limit; the match should hold up to sampling error only if the solves are converged. For the 3D example, rerun with many more LSQR iterations and check whether the pointwise variance maps, especially in poorly observed regions, change materially.

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

If this is right

  • Users can draw independent posterior samples for 3D PAT on general finite-element geometries with roughly 2×10^5 unknowns, with each sample costing one forward and one adjoint wave propagation; the paper reports 100 samples in about 5 hours on 10 MPI processes.
  • Because the backward recursion implements the exact transpose of the discrete forward operator to machine precision, LSQR can be used; the construction also handles non-symmetric discretizations, since the transpose of a sparse matrix costs about the same as the matrix itself.
  • Discretization choices are part of the uncertainty budget: mass lumping cuts the forward-solve time by roughly an order of magnitude with no discernible bias in posterior variance, absorbing boundary conditions must be placed far enough that reflections cannot return during the observation interval, and a Whittle–Matérn prior gives dimension-independent uncertainty behavior that the independent Ga
  • The same matrix-free RTO machinery transfers to any linear PDE-constrained inverse problem with a Gaussian posterior; only the forward and adjoint action are needed.

Where Pith is reading between the lines

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

  • (Editorial extension) A principled LSQR stopping rule that splits the parameter space into data-informed and prior-dominated subspaces—flagged by the paper as open—would likely remove the remaining bias in the variance maps and make RTO posterior samples exact in practice, not just in theory.
  • (Editorial extension) The observed artifacts with the independent Gaussian prior suggest a practical rule for choosing priors: the prior covariance should be dimension-independent or have rapidly decaying eigenvalues; otherwise the finite-element mesh sets the uncertainty scale.
  • (Editorial extension) The enlarged-domain absorbing-boundary strategy could be replaced by an adjoint-consistent perfectly matched layer, which the paper found introduced artifacts; if such a layer can be made exactly transpose-consistent, the domain padding and its extra cost could be avoided.
  • (Editorial extension) For nonlinear extensions such as quantitative PAT with unknown optical parameters, the same RTO construction could be applied after linearizing around the MAP estimate, though the equivalence to exact posterior sampling would then be only approximate.

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 / 6 minor

Summary. The paper presents a matrix-free randomized-then-optimize (RTO) sampler for Bayesian photoacoustic tomography with finite-element discretizations. It constructs a discrete adjoint that is the exact transpose of the discrete forward operator and is consistent with the continuous adjoint; each posterior sample is obtained by solving a randomly perturbed least-squares problem with LSQR using only forward and adjoint wave propagations. Appendix A shows that, for exact solves, the perturbed least-squares solution has the exact posterior mean and covariance. The method is validated against explicitly assembled posterior covariances on a small 2D problem, compared with NUTS, and demonstrated on a 3D problem with about 2e5 unknowns.

Significance. If the numerical implementation delivers exact posterior samples, this is a substantial step: it would be the first large-scale finite-element Bayesian PAT uncertainty quantification, and the matrix-free RTO framework extends to other linear PDE-constrained inverse problems. The paper ships reproducible code, a machine-precision adjoint test, and a clean derivation of RTO equivalence. However, the experimental evidence for posterior variance is weakened by the fixed LSQR iteration limits and the Whittle-Matérn sampling inconsistency, so the main 3D variance maps are not yet validated.

major comments (3)
  1. [Algorithm 5.1, §6.1, §6.3, §6.4] The RTO equivalence in Appendix A (Eq. A.4) is for exact least-squares solves, but Algorithm 5.1 terminates LSQR at a hand-fixed iteration limit (300 in 2D, 350 in 3D, §6.1). The paper itself states that discrepancy-principle stopping in 2–10 iterations is 'insufficient for reliable uncertainty quantification', that 'a stopping criterion for LSQR that accounts for both the prior and the likelihood is needed to accurately recover its magnitude' (§6.3), and that the anomalous low 3D independent-prior standard deviation far from the measurement surface is plausibly due to 'early termination of the LSQR iterations' (§6.4). Consequently, the reported posterior standard deviations are not demonstrated to be draws from N(μ_post, C_post); the main 3D variance maps are potentially biased by Krylov truncation. This is fixable but requires either a proper stopping criterion or numerical evidence of
  2. [§4.1.2 and §6.4] Whittle–Matérn prior inconsistency: §4.1.2 defines C = σ² Z^{-1} (ℓ^{-2}I − Δ)^{-(ν+d/2)}, so for ν=1.5, d=3 the covariance is (ℓ^{-2}I − Δ)^{-3}. Section 6.4 generates samples by solving (ℓ^{-2}I − Δ)^3 p0 = W, which gives p0 with covariance (ℓ^{-2}I − Δ)^{-6}. Thus the 3D Whittle–Matérn posterior is computed for a different prior than the one defined and used in the likelihood. The definition or the sampling procedure must be corrected (e.g., using a fractional power), otherwise the second row of Figure 9 does not quantify uncertainty for the stated prior.
  3. [§6.3] The NUTS comparison does not validate RTO variance magnitudes: NUTS has effective sample size 91 out of 1200 draws and the posterior standard deviations differ in magnitude from RTO; the paper attributes this to LSQR stopping. The small-scale reference in §6.1 is only 2D with 20×20 interior points and agreement is qualitative. To support the claim that RTO produces correct posterior variances, a quantitative comparison with an explicitly computed posterior covariance should be reported for a problem that exercises the same LSQR iteration regime as the large-scale experiments.
minor comments (6)
  1. [Abstract/§1] Typos: 'quatntification', 'Sectyion', 'T omography'.
  2. [§3] The text near the permutation discussion contains 'adjointeee'.
  3. [§4.1.2] Inconsistent spelling 'Whittle-Mat'ern' vs 'Matérn'; please standardize.
  4. [§6.4] After Figure 8, the text contains a duplicated fragment 'inside the inclusion.ion.'
  5. [Figure 6] Panel labels in text and caption are confusing: the text refers to panels (b)/(e) and (c)/(f) for two NUTS variants; please clarify which panel corresponds to which NUTS configuration.
  6. [§4.2] The noise normalization in (2.4) makes the synthetic noise non-Gaussian, as the paper acknowledges. For validation purposes, consider using exact Gaussian noise with fixed σ and no normalization to avoid a mismatch between the data-generating model and the likelihood.

Circularity Check

0 steps flagged

No circularity: the RTO posterior-sample property is proven in Appendix A from stated linear-Gaussian assumptions; numerical validation is self-consistency checking and self-citations are contextual. LSQR fixed-iteration stopping is a numerical limitation, not a circular reduction.

full rationale

The paper's central derivation is self-contained. Appendix A explicitly computes the mean and covariance of the perturbed least-squares solution and shows, in Eqs. A.2–A.4, that it equals the posterior mean and covariance; this is a proof from the stated linear-Gaussian model and does not assume the conclusion. The matrix-free transpose construction in Section 5.1 is verified by the discrete adjoint test, which is an implementation check, not circularity. The comparison of sample-based posterior standard deviation to an explicitly assembled reference covariance on a small mesh is a self-consistency verification: both use the same discrete forward model, so it validates the sampler's internal consistency rather than importing the result as an input. Self-citations to earlier PAT Bayesian work (e.g., [24]–[27]) are contextual comparisons and do not carry the load-bearing argument; no uniqueness theorem from the authors is invoked. The main weakness is numerical rather than circular: Algorithm 5.1 solves each perturbed problem with a fixed LSQR iteration limit (300 in 2D, 350 in 3D, §6.1), and the paper itself states that discrepancy-principle stopping (2–10 iterations) is 'insufficient for reliable uncertainty quantification' and that 'a stopping criterion for LSQR that accounts for both the prior and the likelihood is needed to accurately recover its magnitude' (§6.3), and attributes the 3D independent-prior artifact plausibly to 'early termination of the LSQR iterations' (§6.4). This means the empirical samples may not be exact posterior samples, but that is an approximation/convergence gap, not a circular definition or a fitted-input-called-prediction. A separate modeling inconsistency exists in §6.4: solving (ℓ^{-2}I−∆)^3 p0 = W yields covariance (ℓ^{-2}I−∆)^{-6}, whereas (4.2) with s+d/2 = 3 specifies exponent −3; this is a statistical modeling issue, not circularity. Overall, no step in the derivation reduces to its own inputs, so the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 7 axioms · 0 invented entities

No new physical entities are introduced. The load-bearing hand-tuned inputs are the LSQR stopping limit, domain extensions, and prior hyperparameters; the core mathematics relies on standard linear-Gaussian inverse problem and FEM assumptions.

free parameters (4)
  • LSQR iteration limit = 300 (2D), 350 (3D)
    Hand-chosen fixed stopping limit in §6.1/§6.4; no convergence criterion for posterior samples; paper admits it may stop too early or too late and that posterior std magnitude depends on it.
  • Computational domain extension = 0.75 units (2D), 1 unit radial (3D)
    Chosen ad hoc to keep artificial boundary reflections outside the physical domain during [0,T]; underpins the reconciliation of discrete transpose with continuous adjoint in §5.1.
  • Whittle-Matérn hyperparameters = ℓ=0.15, s=1.5, σ=3.0
    User-specified prior parameters in §6.4; they determine the effective prior smoothness and hence the posterior.
  • Time step Δt = 0.0012, 0.001, 0.0097
    Chosen per experiment to satisfy CFL; temporal discretization is explicitly studied as a factor affecting posterior uncertainty.
axioms (7)
  • domain assumption PAT acoustic propagation is governed by the linear wave equation with constant speed c=1 under thermal and stress confinement.
    §2; standard physical modeling assumption imported from photoacoustic imaging literature, not derived here.
  • domain assumption First-order absorbing boundary condition (2.2) plus a sufficiently enlarged domain approximates the unbounded-domain radiation condition without contaminating the physical region during [0,T].
    §2 and §5.1; used to justify domain truncation and the claim that the algebraic-transpose/continuous-adjoint discrepancy is confined outside the region of interest.
  • domain assumption Ultrasound sensors coincide with finite-element nodes, so the observation operator Π is row selection.
    §3; simplifies the transpose to node injection; paper notes generalization to interpolation transposes is possible.
  • domain assumption Additive Gaussian noise with known absolute scale σ; synthetic noise normalized to exact relative level d_noise is treated as Gaussian in the likelihood.
    §2, §4.2; the paper itself notes the normalized synthetic perturbation is not strictly Gaussian.
  • domain assumption Mass-lumped FEM (M_L) is an adequate approximation of consistent FEM for the wave forward model at used resolutions.
    §3, §6.1; error is non-negligible but decreases with mesh refinement; authors claim no systematic posterior bias.
  • standard math Whittle-Matérn SPDE sampling relation: solving (κ² I - Δ)^α p0 = W with α=(ν+d/2)/2 gives covariance (4.2).
    §4.1.2/§6.4; standard SPDE representation of Matérn fields. The paper states the SPDE with exponent 3, which for s=1.5, d=3 would be the covariance exponent, not the square-root operator exponent; this appears to be a mismatch.
  • standard math The posterior is exactly Gaussian because the forward operator is linear and prior/likelihood are Gaussian; RTO perturbed least-squares solutions have the posterior covariance.
    §4, §5, Appendix A; derivation is self-contained and correct.

pith-pipeline@v1.3.0-alltime-deepseek · 20735 in / 15567 out tokens · 142659 ms · 2026-08-01T07:37:52.730236+00:00 · methodology

0 comments
read the original abstract

Photoacoustic tomography (PAT) is a promising modality for high-resolution biomedical imaging, motivating the need for reliable uncertainty quantification (UQ) of reconstructed images. Bayesian approaches provide a rigorous framework for UQ but remain computationally challenging for realistic three-dimensional PAT and are sensitive to numerical approximations in the governing wave equation. We develop a finite-element Bayesian UQ framework for PAT that accommodates complex computational domains and detector geometries while enabling large-scale three-dimensional inference. The proposed methodology reformulates the randomize-then-optimize (RTO) sampling strategy as a matrix-free algorithm that generates independent posterior samples using only forward and adjoint wave propagations. Particular attention is given to constructing an adjoint discretization that forms an exact transpose pair with the discrete forward operator while remaining consistent with the continuous PAT adjoint, enabling efficient least-squares solvers within the sampling procedure. We investigate the influence of temporal discretization, artificial boundary conditions, and adjoint consistency on posterior uncertainty and identify discretization strategies that avoid numerical artifacts. The framework is validated against exact posterior statistics, existing Bayesian PAT methods, and Hamiltonian Monte Carlo using the No-U-Turn Sampler (NUTS), and is demonstrated on a three-dimensional problem with approximately $2\times 10^5$ unknowns on a general finite-element domain. To the best of our knowledge, this is the first large-scale Bayesian PAT study on general three-dimensional finite-element geometries, and the methodology extends naturally to a broad class of linear PDE-constrained inverse problems.

Figures

Figures reproduced from arXiv: 2607.21373 by Amal Mohammed A Alghami, Babak Maboudi Afkham, Hassan Yazdanian, Tanja Tarvainen.

Figure 1
Figure 1. Figure 1: Reference pointwise posterior variance computed from the explicitly assembled [PITH_FULL_IMAGE:figures/full_fig_p014_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Relative forward error (left vertical axis) and computation time (right vertical [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: LSQR convergence diagnostics for the two- and three-dimensional test cases. [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: True initial pressures and corresponding boundary measurements for the 2D [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Empirical mean and pointwise posterior standard deviation estimated from [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Empirical mean and pointwise posterior standard deviation estimated from [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: True phantom (only the inclusion surface) and computational geometry for [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Mean/MAP estimates for the posterior constructed with (a) the independent [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Point-wise standard deviation. The first and second rows correspond to the [PITH_FULL_IMAGE:figures/full_fig_p022_9.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

55 extracted references

  1. [1]

    S. R. Arridge and M. M. Betcke and B. T. Cox and F. Lucka and B. E. Treeby , title =. 2016 , OPTkey =

    2016

  2. [2]

    C. E. Rasmussen and C. K. I. Williams , title=

  3. [3]

    The Journal of the Acoustical Society of America , volume=

    Fast calculation of pulsed photoacoustic fields in fluids using k-space methods , author=. The Journal of the Acoustical Society of America , volume=. 2005 , publisher=

    2005

  4. [4]

    Kaipio and E

    J. Kaipio and E. Somersalo , ALTeditor =. Statistical and Computational Inverse Problems , publisher =. 2005 , OPTkey =

    2005

  5. [5]

    SIAM Journal on Imaging Sciences , volume=

    Utilizing variational autoencoders in the Bayesian inverse problem of photoacoustic tomography , author=. SIAM Journal on Imaging Sciences , volume=. 2023 , publisher=

    2023

  6. [6]

    doi:, journal =

    Solving Bayesian Inverse Problems via Variational Autoencoders , author =. doi:, journal =

  7. [7]

    Biomedical physics & engineering express , volume=

    Modelling of errors due to speed of sound variations in photoacoustic tomography using a Bayesian framework , author=. Biomedical physics & engineering express , volume=. 2019 , publisher=

    2019

  8. [8]

    Journal of Biomedical Optics , volume=

    Deep learning in photoacoustic tomography: current approaches and future directions , author=. Journal of Biomedical Optics , volume=

  9. [9]

    Tick and A

    J. Tick and A. Pulkkinen and T. Tarvainen , title =. 2016 , OPTkey =

    2016

  10. [10]

    Pulkkinen and B

    A. Pulkkinen and B. T. Cox and S. R. Arridge and J. P. Kaipio and T. Tarvainen , title =. 2014 , OPTkey =

    2014

  11. [11]

    Beard , title =

    P. Beard , title =. Interface Focus , year =

  12. [12]

    Li and L

    C. Li and L. V. Wang , title =. 2009 , OPTkey =

    2009

  13. [13]

    doi:10.1137/140951758 , journal =

    An Adaptive Shifted Power Method for Computing Generalized Tensor Eigenpairs , author =. doi:10.1137/140951758 , journal =

    doi:10.1137/140951758

  14. [14]

    Nick Higham , title =

  15. [15]

    Kolda and Ali Pinar , eprint =

    Chengbin Peng and Tamara G. Kolda and Ali Pinar , eprint =. Accelerating Community Detection by Using

  16. [16]

    and Zhang, Shanrong and Merritt, Matthew E

    Woessner, Donald E. and Zhang, Shanrong and Merritt, Matthew E. and Sherry, A. Dean , title =. Magnetic Resonance in Medicine , doi =

  17. [17]

    2003 , eid =

    Properties of Highly Clustered Networks , author =. 2003 , eid =. doi:10.1103/PhysRevE.68.026121 , journal =

    doi:10.1103/physreve.68.026121 2003

  18. [18]

    Clawpack Software , author =

  19. [19]

    : A Document Preparation System

    Leslie Lamport. : A Document Preparation System. 1986

    1986

  20. [20]

    Frank Mittlebach and Michel Goossens , title =

  21. [21]

    and Van Loan, Charles F

    Golub, Gene H. and Van Loan, Charles F. , title =

  22. [22]

    Paul Dawkins , title =

  23. [23]

    User's Guide for the

  24. [24]

    Michael Downes , title =

  25. [25]

    Christian Feuers\"anger , title =

  26. [26]

    Oberwolfach Reports , volume=

    Geometric numerical integration , author=. Oberwolfach Reports , volume=

  27. [27]

    SIAM Journal on Scientific Computing , volume=

    Randomize-then-optimize: A method for sampling from posterior distributions in nonlinear inverse problems , author=. SIAM Journal on Scientific Computing , volume=. 2014 , publisher=

    2014

  28. [28]

    1998 , publisher=

    Rank-deficient and discrete ill-posed problems: numerical aspects of linear inversion , author=. 1998 , publisher=

    1998

  29. [29]

    2003 , publisher=

    Iterative methods for sparse linear systems , author=. 2003 , publisher=

    2003

  30. [30]

    2010 , publisher=

    Discrete inverse problems: insight and algorithms , author=. 2010 , publisher=

    2010

  31. [31]

    Medical physics , volume=

    Prospects of photoacoustic tomography , author=. Medical physics , volume=. 2008 , publisher=

    2008

  32. [32]

    Electromagnetic waves (Cambridge, Mass.) , volume=

    Photoacoustic tomography: principles and advances , author=. Electromagnetic waves (Cambridge, Mass.) , volume=. 2014 , publisher=

    2014

  33. [33]

    Journal of Biomedical Optics , volume=

    Quantitative photoacoustic tomography: modeling and inverse problems , author=. Journal of Biomedical Optics , volume=. 2024 , publisher=

    2024

  34. [34]

    Inverse Problems , volume=

    Reconstructing absorption and scattering distributions in quantitative photoacoustic tomography , author=. Inverse Problems , volume=. 2012 , publisher=

    2012

  35. [35]

    IEEE transactions on medical imaging , volume=

    Bayesian image reconstruction in quantitative photoacoustic tomography , author=. IEEE transactions on medical imaging , volume=. 2013 , publisher=

    2013

  36. [36]

    Characterizing Impacts of Model Uncertainties in Quantitative Photoacoustics , journal =

    Ren, Kui and Vall\'. Characterizing Impacts of Model Uncertainties in Quantitative Photoacoustics , journal =. 2020 , doi =

    2020

  37. [37]

    Data-driven Models in Inverse Problems , volume=

    Utilizing uncertainty quantification variational autoencoders in inverse problems with applications in photoacoustic tomography , author=. Data-driven Models in Inverse Problems , volume=. 2024 , publisher=

    2024

  38. [38]

    The Journal of the Acoustical Society of America , volume=

    Three dimensional photoacoustic tomography in Bayesian framework , author=. The Journal of the Acoustical Society of America , volume=. 2018 , publisher=

    2018

  39. [39]

    Photons Plus Ultrasound: Imaging and Sensing 2016 , volume=

    Bayesian parameter estimation in spectral quantitative photoacoustic tomography , author=. Photons Plus Ultrasound: Imaging and Sensing 2016 , volume=. 2016 , organization=

    2016

  40. [40]

    IEEE Transactions on Medical Imaging , volume=

    Modeling of errors due to uncertainties in ultrasound sensor locations in photoacoustic tomography , author=. IEEE Transactions on Medical Imaging , volume=. 2020 , publisher=

    2020

  41. [41]

    WHITTLE-MAT

    Roininen, Lassi and Huttunen, Janne MJ and Lasanen, Sari , journal=. WHITTLE-MAT

  42. [42]

    2012 , publisher=

    Gaussian random processes , author=. 2012 , publisher=

    2012

  43. [43]

    2011 , publisher=

    Numerical methods for large eigenvalue problems: revised edition , author=. 2011 , publisher=

    2011

  44. [44]

    2014 , publisher=

    Numerical methods for partial differential equations , author=. 2014 , publisher=

    2014

  45. [45]

    Communications on pure and applied mathematics , volume=

    Radiation boundary conditions for acoustic and elastic wave calculations , author=. Communications on pure and applied mathematics , volume=

  46. [46]

    2002 , publisher=

    The finite element method for elliptic problems , author=. 2002 , publisher=

    2002

  47. [47]

    2008 , publisher=

    The mathematical theory of finite element methods , author=. 2008 , publisher=

    2008

  48. [48]

    SIAM Journal on Scientific Computing , volume=

    Bayesian inverse problems with l\_1 priors: a randomize-then-optimize approach , author=. SIAM Journal on Scientific Computing , volume=. 2017 , publisher=

    2017

  49. [49]

    ACM Transactions on Mathematical Software (TOMS) , volume=

    LSQR: An algorithm for sparse linear equations and sparse least squares , author=. ACM Transactions on Mathematical Software (TOMS) , volume=. 1982 , publisher=

    1982

  50. [50]

    Photons Plus Ultrasound: Imaging and Sensing 2025 , volume=

    Exploring uncertainty quantification for photoacoustic image reconstruction and quantitative oxygenation mapping , author=. Photons Plus Ultrasound: Imaging and Sensing 2025 , volume=. 2025 , organization=

    2025

  51. [51]

    Bildverarbeitung f

    Invertible neural networks for uncertainty quantification in photoacoustic imaging , author=. Bildverarbeitung f. 2021 , organization=

    2021

  52. [52]

    SIAM Journal on Scientific Computing , volume=

    Fixing nonconvergence of algebraic iterative reconstruction with an unmatched backprojector , author=. SIAM Journal on Scientific Computing , volume=. 2019 , publisher=

    2019

  53. [53]

    International Conference on Scale Space and Variational Methods in Computer Vision , pages=

    A Novel Interpretation of the Radon Transform’s Ray and Pixel-Driven Discretizations Under Balanced Resolutions , author=. International Conference on Scale Space and Variational Methods in Computer Vision , pages=. 2025 , organization=

    2025

  54. [54]

    International Journal of Solids and Structures , volume=

    Finite element mass matrix lumping by numerical integration with no convergence rate loss , author=. International Journal of Solids and Structures , volume=. 1975 , publisher=

    1975

  55. [55]

    and Yazdanian, Hassan and Tarvainen, Tanja , title =

    Afkham, Babak Maboudi and Alghamdi, Amal Mohammed A. and Yazdanian, Hassan and Tarvainen, Tanja , title =. 2026 , howpublished =

    2026