Pith. sign in

REVIEW 4 major objections 4 minor 34 references

Statistical Reconstruction For Anisotropic X-ray Dark-Field Tomography

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Statistical reconstruction for dark-field X-ray tomography becomes practical through a simplified Rician-noise model that matches the full likelihood's quality at roughly one-eighteenth the compute time.

desk verdict Plausible m2 implementation and useful experiments, but the new m3 model has an internally inconsistent printed gradient and no code—core claim unverifiable. read the letter →

arxiv 2501.03160 v1 pith:XETNE5HF submitted 2025-01-06 cs.CE

David Frank , Cederik Höfs , Tobias Lasser This is my paper · ORCID
classification cs.CE
keywords AnisotropicX-rayDark-FieldTomographyStatisticalreconstructionRiciannoiseGratinginterferometryLipschitzconstantFastgradientmethodFiberorientationextractionCT
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that statistically sound reconstruction for Anisotropic X-ray Dark-Field Tomography (AXDT) is both feasible and worth the extra cost: it provides a numerically stable implementation of the Rician-likelihood model m2, and it introduces a simplified model m3 that keeps the same noise assumptions while treating attenuation as known. On two experimental datasets — crossed wooden sticks and a human cerebellum — the statistical models m2 and m3 produce visibly cleaner scattering maps and fiber orientations than the linearized model m1 used in all previous AXDT work. The key practical claim is that m3 matches m2's reconstruction quality while cutting wall-clock time from about eighteen hours to about one hour on the sticks dataset, with m1 fastest but noisiest. The paper also derives Lipschitz bounds for m1 and m3, giving first-order optimizers a principled step size. If these claims hold, AXDT's noise performance improves enough to make fiber-microstructure imaging practical at lower X-ray dose.

What carries the argument

The machinery is the Rician amplitude noise model of grating interferometry combined with the discretized AXDT forward operator B. After phase stepping, the measured intensity coefficients obey Aj ∼ N(aj, aj/N) for the mean and (1/2)Bj ∼ Rician(bj/2, sqrt(aj/(2N))) for the visibility half-amplitude, and the reconstruction problems are negative log-likelihoods built from these distributions. The new model m3 is the same Rician likelihood with the attenuation term removed: it optimizes only the spherical-harmonic scattering coefficients η via the forward operator B, the dark-field signal d_η = exp(Bη), and the reference visibility α. Two numerical-stability choices carry the implementation: evaluating ln I0 and the ratio I1/I0 directly rather than the modified Bessel functions themselves, which would overflow. The final piece is the Lipschitz bound Lm3 ≤ N max(aα² + bα) (K/(4π)||A||)², which lets fast gradient methods choose step sizes without line search.

What would settle it

Acquire a phantom with known attenuation and scattering directions, then reconstruct with m1, m2, and m3, feeding m3 an attenuation map corrupted by realistic noise. If m3's scattering coefficients degrade toward m1's quality while m2 stays accurate, the central assumption of m3 fails. A simpler check: fit the amplitude pairs from a grating-interferometry scan to the predicted Rician distribution and test whether the empirical variance matches aj/N.

Watch

Extended reading notes

Core claim

The central discovery is that the full joint statistical model for AXDT can be simplified without giving up its noise model. The paper starts from the grating-interferometry result that Fourier-extracted amplitudes follow a normal distribution for the mean and a Rician distribution for the visibility half-amplitude, and builds m2 as the resulting negative log-likelihood over both attenuation µ and scattering coefficients η. It then derives m3 by assuming the attenuation map is exactly known, producing the loss fm3(η) = Σ [ (N/4) a α² d_η² − ln I0((N/2) b α d_η) ] with d_η = exp(Bη), the forward-projected dark-field signal, and α the reference visibility. Both models are shown to be locally convex, and m3 admits an explicit Lipschitz bound on its gradient, enabling accelerated first-order optimization. In experiments, m2 and m3 improve noise behavior over m1 in the spherical-harmonic coefficients and in the extracted fiber orientations, and m3 reaches comparable quality to m2 in a small fraction of the iterations and time.

Load-bearing premise

The claim that m3 matches m2's quality rests on the assumption that the attenuation map is exactly known and constant, plus the assumption that the Rician distribution correctly describes the noise in the measured Fourier amplitudes; if the attenuation estimate is noisy or the Rician model mismatches the real statistics, the advantage of the statistical models may shrink or disappear.

Editorial extensions

If this is right

  • Practitioners can now run statistically correct AXDT reconstruction on measured data without unstable Bessel-function arithmetic, so the m2 model becomes usable rather than just theoretical.
  • The m3 model makes the noise benefit of statistical reconstruction affordable: a full sticks dataset reconstructs in about an hour instead of eighteen, bringing statistical AXDT into routine workflows.
  • The explicit Lipschitz bound for m3 gives fast gradient methods a provable step size, removing the trial-and-error tuning that currently dominates algorithm choice.
  • Because m3 optimizes only the scattering coefficients, regularization or prior models for fiber structure can be added directly to the loss without touching the joint attenuation estimation.
  • Cleaner scattering maps translate into more reliable fiber-orientation extraction, which is the clinically and materially relevant output of AXDT.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural pipeline the paper does not test is two-step: reconstruct attenuation with standard CT, then run m3 with that attenuation fixed; this would deliver m2-class noise performance without ever running the expensive joint model.
  • The numerical-stability tricks for Bessel functions transfer to any Rician-likelihood reconstruction in grating-based dark-field imaging, so other CT and radiography methods could adopt them.
  • The variance statistics in Table 2 raise the question of whether m3 is actually slightly more stable than m2 on the full volume; a formal efficiency comparison would clarify whether 'identical performance' is exact or just close on these datasets.
  • If the attenuation-known assumption proves too strong, a block-coordinate scheme could alternate between m3 scattering updates and attenuation re-estimation from the same phase-stepping data, an extension the authors do not explore.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes numerically stable implementations of two statistical reconstruction models for Anisotropic X-ray Dark-Field Tomography (AXDT), building on the Rician amplitude noise model from Chabior et al. and the statistical formulation by Schilling et al. The main contributions are (i) a stable implementation of the full statistical model m2, (ii) a new simplified model m3 that treats attenuation as known and constant, (iii) Lipschitz constant bounds for m1 and m3, and (iv) an experimental comparison of m1, m2, and m3 on a crossed-sticks phantom and a human cerebellum dataset, using CG, FGM, and L-BFGS optimizers. The central claim is that m3 achieves essentially identical reconstruction quality to m2 at substantially lower computational cost, while both statistical models outperform the linearized m1 in noise behavior.

Significance. If the claims hold, this work would be a practically valuable contribution to AXDT reconstruction: it would provide a computationally feasible statistical reconstruction method that retains the Rician noise model, plus convergence-relevant Lipschitz bounds. The manuscript also honestly acknowledges that the m2 Lipschitz bound is not yet found, and the empirical setup compares against a previously published baseline on independent datasets. However, the significance is currently undermined by a concrete inconsistency in the m3 gradient formula, the absence of released code or data to verify which objective was actually minimized, and a lack of error bars or ground-truth comparisons for the quality claims. The central 'identical performance' claim is therefore not verifiable from the manuscript as written.

major comments (4)
  1. [§2.2, m3 gradient] The displayed gradient for m3 is not the derivative of the displayed loss. For f_m3(η) = 1^T( N/4 a α^2 d_η^2 − ln I0( N/2 b α d_η ) ), differentiating with respect to Bη gives B^*( N/2 a α^2 d_η^2 − (N/2 b α d_η) I1/I0( (N/2 b α d_η) ) ), but the paper prints ∇f_m3(η) = B^*( z I1(z)/I0(z) − N/2 a α^2 d_η^2 ) with z = N/2 b α d_η^2. The printed z would need to be N/2 b α d_η for consistency with the loss and with the printed Hessian. Since at least two of the three displayed equations for m3 disagree, and no code or data are provided, the reader cannot determine which objective was actually optimized in Figure 2 and Table 2. This is load-bearing because the central claim that m3 reproduces m2's quality at roughly 1/18 of the runtime depends on the correct m3 objective being implemented.
  2. [§2.2, Lipschitz Condition] The Lipschitz bound derivation for m3 is incomplete and appears to contain sign errors. The Hessian diagonal term is written with a minus sign as N a α^2 d_η^2 + z^2[(I1/I0)^2 − 1], but the displayed bound then manipulates |N a α^2 d_η^2 − z^2(1 − I1/I0)|^2, which does not correspond to the diagonal term. Moreover, the step bounding ||B|| as ≤ K/(4π)||A|| omits the sum over m and the factor (2k+1), so the stated bound is not justified as written. These issues affect the FGM step-size recommendations that the paper claims as a contribution.
  3. [Experiments, Fiber extraction] The visual comparison uses post-hoc iteration selection: 'We extracted the fibers at different iterations and chose the visually best results for the final visualization.' This makes the qualitative claims in Section 4 difficult to interpret, and no error bars or quantitative image-quality metrics are provided for the fiber visualizations. The paper should report a fixed, pre-specified stopping criterion for all methods, or justify the selected iterations with quantitative criteria such as loss values or image-quality metrics.
  4. [Experiments, Crossed Sticks Sample] The claim that the statistical models outperform m1 in noise behavior rests on the statistics in Table 2 and visual inspection of Figure 3. However, Table 2 reports only point estimates of mean, variance, and q95 for a single reconstruction per model; no confidence intervals, repeated experiments, or ground-truth comparison are provided. The 'identical performance' of m2 and m3 is also based on single runs. Given the m3 gradient inconsistency and the absence of code/data, these central claims are not quantitatively supported.
minor comments (4)
  1. [Abstract and Introduction] The abstract says 'our newly proposed formulation proves highly efficient with identical performance,' but the manuscript itself only shows that m3 achieves a similar visual appearance and similar summary statistics on two datasets; 'identical' is too strong a claim and should be qualified.
  2. [§2.2, m3 notation] The notation for d_η is inconsistent: sometimes written as dη and sometimes as d^2_η, particularly in the gradient and Hessian formulas. This makes it harder to spot the algebraic error and should be cleaned up.
  3. [§3, Reconstruction] The description says 'For model m2, we choose fixed step lengths in a uniform interval,' but the paper does not state which step lengths were actually used for the FGM results reported in Figure 2, making the FGM comparison incomplete.
  4. [§4, Crossed Sticks Sample] In the discussion of Figure 4, the text says 'For m1 and m3, we ran 280 iterations of CG and L-BFGS respectively, and 1800 iterations of L-BFGS for m2,' but Table 1 reports 240 iterations for m1 and m3. This numerical inconsistency should be resolved.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: m3 is an explicit fixed-attenuation special case of the published m2 model, and the performance comparison is empirical, not a fit; the printed m3 gradient inconsistency is a correctness/verifiability issue, not a circular-input issue.

full rationale

Walking the derivation chain, the paper's inputs are external: the AXDT forward model B is from Wieczorek et al. (2016, ref. [13]) and the Rician amplitude statistics are from Chabior et al. (2011, ref. [29]). Model m2 is the negative log-likelihood derived by Schilling et al. (2017, ref. [17]) from those statistics, and model m3 is explicitly presented as the same Rician assumption with attenuation treated as known and constant. The central claim that m3 matches m2's quality is an empirical comparison on two datasets (Figure 2, Table 2), not a parameter fitted to force that outcome; no m3 parameter is tuned to make Table 2 match m2. The Lipschitz bound for m3 is derived from the printed Hessian, and the local-convexity remark relies on a prior proof for m2, but that citation does not by itself force the experimental conclusion. The notable concern in the manuscript is a correctness/reproducibility problem rather than circularity: in Section 2.2, the printed gradient with z=(N/2)bαdη² is not the derivative of the printed m3 loss, and no code is released, so it is unclear which objective produced Figures 2-5 and Table 2. That is an internal-consistency and verifiability issue, not a reduction of the result to its own inputs. Therefore the circularity score is low.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The central claim does not introduce new physical entities or fitted constants. It rests on previously published noise and forward models, plus the m3 simplification that attenuation is known exactly.

assumptions (5)
  • domain assumption The extracted Fourier amplitudes follow A_j ~ N(a_j, a_j/N) and 1/2 B_j ~ R(b_j/2, sqrt(a_j/(2N))).
    Inherited from Chabior et al. [29] and the basis for the statistical models m2 and m3; if this noise model is incorrect, the likelihood functions are misspecified.
  • domain assumption The AXDT forward model B and its discretization with spherical harmonics (truncation K=4, even degrees) accurately describe dark-field scattering.
    From Wieczorek et al. [13, 14]; the reconstruction quality claims depend on this forward model.
  • domain assumption Physics dictate a < a_mu (A_mu >= 0) and d <= 1 (B_eta >= 0), so the domain can be restricted accordingly.
    Used to restrict the domain for the Lipschitz bound derivation.
  • ad hoc to paper The attenuation a is known exactly and can be treated as constant in m3.
    This simplification defines m3; if the measured attenuation is noisy or biased, m3's reconstruction quality degrades.
  • ad hoc to paper Model m2 is locally convex (from Schilling et al. 2017), and m3 is similarly locally convex.
    Asserted as straightforward without proof; the Hessian of m3 has negative entries, so convexity holds only on a restricted, unspecified region.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Statistical Reconstruction For Anisotropic X-ray Dark-Field Tomography." pith.science (2026). https://pith.science/paper/XETNE5HF

@misc{pith2026250103160,
  author       = {Pith},
  title        = {Pith review of: Statistical Reconstruction For Anisotropic X-ray Dark-Field Tomography},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XETNE5HF}},
  note         = {Machine review of arXiv:2501.03160}
}
read the original abstract

Anisotropic X-ray Dark-Field Tomography (AXDT) is a novel imaging technology that enables the extraction of fiber structures on the micrometer scale, far smaller than standard X-ray Computed Tomography (CT) setups. Directional and structural information is relevant in medical diagnostics and material testing. Compared to existing solutions, AXDT could prove a viable alternative. Reconstruction methods in AXDT have so far been driven by practicality. Improved methods could make AXDT more accessible. We contribute numerically stable implementations and validation of advanced statistical reconstruction methods that incorporate the statistical noise behavior of the imaging system. We further provide a new statistical reconstruction formulation that retains the advanced noise assumptions of the imaging setup while being efficient and easy to optimize. Finally, we provide a detailed analysis of the optimization behavior for all models regarding AXDT. Our experiments show that statistical reconstruction outperforms the previously used model, and particularly the noise performance is superior. While the previously proposed statistical method is effective, it is computationally expensive, and our newly proposed formulation proves highly efficient with identical performance. Our theoretical analysis opens the possibility to new and more advanced reconstruction algorithms, which in turn enable future research in AXDT.

Figures

Figures reproduced from arXiv: 2501.03160 by the authors.

Figure 1
Figure 1. Schematic overview of grating-based interferometry setup used in Anisotropic X-ray Dark-Field Tomography. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Fiber extraction For fiber extraction, we chose the reconstruction from L-BFGS for the nonlinear models. For model m1, we chose CG, even though L-BFGS converged slightly faster; previous work used CG, and as such, we want to compare our research to the previously used algorithm. We verified that the reconstructions with similar losses have a small mean square error. We extracted the fibers at different iterations an… view at source ↗
Figure 2
Figure 2. Convergence plot for reconstruction of the crossed sticks sample. Top: plot of loss over number of iterations. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: 4 Discussion Convergence Analysis The convergence results from our experiments can be seen in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 3
Figure 3. Figure 3: Visualization of slice 68 of the spherical harmonic coefficients [PITH_FULL_IMAGE:figures/full_fig_p008_3.png]
Figure 4
Figure 4. Figure 4: Fiber visualization of two slices from the crossed wooden sticks sample for each of the models overlaid on top [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Fiber visualization of the center slice from the human brain sample for each of the models overlaid on top of [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 30 canonical work pages

  1. [1]

    Inferring microstructural features and the physiological state of tissues from diffusion-weighted images,

    P. J. Basser, “Inferring microstructural features and the physiological state of tissues from diffusion-weighted images,” NMR in Biomedicine, vol. 8, no. 7, pp. 333–344, 1995

    work page 1995

  2. [2]

    Diffusion Tensor Imaging (DTI)-based White Matter Mapping in Brain Research: A Review,

    Y . Assaf and O. Pasternak, “Diffusion Tensor Imaging (DTI)-based White Matter Mapping in Brain Research: A Review,” Journal of Molecular Neuroscience, vol. 34, no. 1, pp. 51–61, Jan. 2008

    work page 2008

  3. [3]

    Grating-based x-ray dark-field CT for lung cancer diagnosis in mice,

    P. Guo, L. Zhang, J. Lu, H. Zhang, X. Zhu, C. Wu, X. Zhan, H. Yin, Z. Wang, Y . Xu, and Z. Wang, “Grating-based x-ray dark-field CT for lung cancer diagnosis in mice,” European Radiology Experimental, vol. 8, no. 1, p. 12, Jan

    work page

  4. [4]

    Dark-field chest X-ray imaging for the assessment of COVID -19-pneumonia,

    M. Frank, F. T. Gassert, T. Urban, K. Willer, W. Noichl, R. Schick, M. Schultheiss, M. Viermetz, B. Gleich, F. De Marco, J. Herzen, T. Koehler, K. J. Engel, B. Renger, F. G. Gassert, A. Sauter, A. A. Fingerle, B. Haller, M. R. Makowski, D. Pfeiffer, and F. Pfeiffer, “Dark-field chest X-ray imaging for the assessment of COVID -19-pneumonia,” Communications...

    work page 2022

  5. [5]

    X-ray Dark-field Radiography - In-Vivo Diagnosis of Lung Cancer in Mice,

    K. Scherer, A. Yaroshenko, D. A. Bölükbas, L. B. Gromann, K. Hellbach, F. G. Meinel, M. Braunagel, J. von Berg, O. Eickelberg, M. F. Reiser, F. Pfeiffer, S. Meiners, and J. Herzen, “X-ray Dark-field Radiography - In-Vivo Diagnosis of Lung Cancer in Mice,” Scientific Reports, vol. 7, no. 1, p. 402, Mar. 2017

    work page 2017

  6. [6]

    Small angle x-ray scattering with edge-illumination,

    P. Modregger, T. P. Cremona, C. Benarafa, J. C. Schittny, A. Olivo, and M. Endrizzi, “Small angle x-ray scattering with edge-illumination,” Scientific Reports, vol. 6, no. 1, p. 30940, Aug. 2016

    work page 2016

  7. [7]

    Improved In vivo Assessment of Pulmonary Fibrosis in Mice using X-Ray Dark-Field Radiography,

    A. Yaroshenko, K. Hellbach, A. Ö. Yildirim, T. M. Conlon, I. E. Fernandez, M. Bech, A. Velroyen, F. G. Meinel, S. Auweter, M. Reiser, O. Eickelberg, and F. Pfeiffer, “Improved In vivo Assessment of Pulmonary Fibrosis in Mice using X-Ray Dark-Field Radiography,” Scientific Reports, vol. 5, p. 17492, Dec. 2015

    work page 2015

  8. [8]

    Grating-based X-ray Dark-field Computed Tomography of Living Mice,

    A. Velroyen, A. Yaroshenko, D. Hahn, A. Fehringer, A. Tapfer, M. Müller, P. B. Noël, B. Pauwels, A. Sasov, A. Ö. Yildirim, O. Eickelberg, K. Hellbach, S. D. Auweter, F. G. Meinel, M. F. Reiser, M. Bech, and F. Pfeiffer, “Grating-based X-ray Dark-field Computed Tomography of Living Mice,” EBioMedicine, vol. 2, no. 10, pp. 1500–1506, Oct. 2015

    work page 2015

Show all 34 references
  1. [9]

    X-ray phase imaging with a grating interferometer,

    T. Weitkamp, A. Diaz, C. David, F. Pfeiffer, M. Stampanoni, P. Cloetens, and E. Ziegler, “X-ray phase imaging with a grating interferometer,” Optics Express, vol. 13, no. 16, pp. 6296–6304, Aug. 2005, publisher: Optica Publishing Group. [Online]. Available: https://opg.optica....

    2005

  2. [10]

    Phase retrieval and differential phase-contrast imaging with low-brilliance X-ray sources,

    F. Pfeiffer, T. Weitkamp, O. Bunk, and C. David, “Phase retrieval and differential phase-contrast imaging with low-brilliance X-ray sources,” Nature Physics, vol. 2, no. 4, pp. 258–261, Apr. 2006, publisher: Nature Publishing Group. [Online]. Available: https://www.nature.com/...

    2006

  3. [11]

    Hard-X-ray dark-field imaging using a grating interferometer,

    F. Pfeiffer, M. Bech, O. Bunk, P. Kraft, E. F. Eikenberry, C. Brönnimann, C. Grünzweig, and C. David, “Hard-X-ray dark-field imaging using a grating interferometer,” Nature Materials, vol. 7, no. 2, pp. 134–137, Feb. 2008, publisher: Nature Publishing Group. [Online]. Availabl...

    2008

  4. [12]

    Dark-field computed tomography reaches the human scale,

    M. Viermetz, N. Gustschin, C. Schmid, J. Haeusele, M. von Teuffenbach, P. Meyer, F. Bergner, T. Lasser, R. Proksa, T. Koehler, and F. Pfeiffer, “Dark-field computed tomography reaches the human scale,”Proceedings of the National Academy of Sciences, vol. 119, no. 8, p. e211879...

    2022 doi

  5. [13]

    Anisotropic X-Ray Dark-Field Tomography: A Continuous Model and its Discretization,

    M. Wieczorek, F. Schaff, F. Pfeiffer, and T. Lasser, “Anisotropic X-Ray Dark-Field Tomography: A Continuous Model and its Discretization,” Physical Review Letters, vol. 117, no. 15, p. 158101, Oct. 2016. [Online]. Available: https://link.aps.org/doi/10.1103/PhysRevLett.117.158101

    2016 doi

  6. [14]

    Micro-structure orientation extraction for Anisotropic X-Ray Dark-Field Tomography,

    M. Wieczorek, F. Pfeiffer, and T. Lasser, “Micro-structure orientation extraction for Anisotropic X-Ray Dark-Field Tomography,” inProc. Int’l Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine (Fully3D), 2017

    2017

  7. [15]

    Brain Connectivity Exposed by Anisotropic X-ray Dark -field Tomography,

    M. Wieczorek, F. Schaff, C. Jud, D. Pfeiffer, F. Pfeiffer, and T. Lasser, “Brain Connectivity Exposed by Anisotropic X-ray Dark -field Tomography,” Scientific Reports, vol. 8, no. 1, p. 14345, Sep. 2018, number: 1 Publisher: Nature Publishing Group. [Online]. Available: https:...

    2018

  8. [16]

    Statistical image reconstruction methods for transmission tomography,

    J. A. Fessler, “Statistical image reconstruction methods for transmission tomography,” inHandbook of Medical Imaging, Volume 2. Medical Image Processing and Analysis, M. Sonka and J. M. Fitzpatrick, Eds. 1000 20th Street, Bellingham, W A 98227-0010 USA: SPIE, Jun. 2000, ch. 10

    2000

  9. [17]

    Statistical Models for Anisotropic X-Ray Dark- field Tomography,

    N. Schilling, M. Wieczorek, and T. Lasser, “Statistical Models for Anisotropic X-Ray Dark- field Tomography,” Proc. International Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine (Fully3D) , pp. 677–680, Jun. 2017. [Online]. Available: ...

    2017 doi

  10. [18]

    Optimized first-order methods for smooth convex minimization,

    D. Kim and J. A. Fessler, “Optimized first-order methods for smooth convex minimization,” Mathematical Programming, vol. 159, no. 1, pp. 81–107, Sep. 2016

    2016

  11. [19]

    Six dimensional X-ray Tensor Tomography with a compact laboratory setup,

    Y . Sharma, M. Wieczorek, F. Schaff, S. Seyyedi, F. Prade, F. Pfeiffer, and T. Lasser, “Six dimensional X-ray Tensor Tomography with a compact laboratory setup,”Applied Physics Letters, vol. 109, no. 13, p. 134102, Sep. 2016

    2016

  12. [20]

    Advanced acquisition methods for anisotropic X-ray dark-field imaging,

    Y . Sharma, “Advanced acquisition methods for anisotropic X-ray dark-field imaging,” phd, Technische Universität München, 2018

    2018

  13. [21]

    The Convergence of a Class of Double-rank Minimization Algorithms 1. General Considerations,

    C. G. Broyden, “The Convergence of a Class of Double-rank Minimization Algorithms 1. General Considerations,” IMA Journal of Applied Mathematics, vol. 6, no. 1, pp. 76–90, 1970

    1970

  14. [22]

    A new approach to variable metric algorithms,

    R. Fletcher, “A new approach to variable metric algorithms,”The Computer Journal, vol. 13, no. 3, pp. 317–322, Jan. 1970

    1970

  15. [23]

    A family of variable-metric methods derived by variational means,

    D. Goldfarb, “A family of variable-metric methods derived by variational means,”Mathematics of Computation, vol. 24, no. 109, pp. 23–26, 1970

    1970

  16. [24]

    Conditioning of quasi-Newton methods for function minimization,

    D. F. Shanno, “Conditioning of quasi-Newton methods for function minimization,”Mathematics of Computation, vol. 24, no. 111, pp. 647–656, 1970

    1970

  17. [25]

    On the limited memory bfgs method for large scale optimization,

    D. C. Liu and J. Nocedal, “On the limited memory bfgs method for large scale optimization,” Mathematical programming, vol. 45, no. 1, pp. 503–528, 1989

    1989

  18. [26]

    Methods of conjugate gradients for solving linear systems,

    M. Hestenes and E. Stiefel, “Methods of conjugate gradients for solving linear systems,” Journal of Research of the National Bureau of Standards , vol. 49, no. 6, p. 409, Dec. 1952. [Online]. Available: http://dx.doi.org/10.6028/jres.049.044

    1952 doi

  19. [27]

    An introduction to the conjugate gradient method without the agonizing pain,

    J. R. Shewchuk et al., “An introduction to the conjugate gradient method without the agonizing pain,” 1994

    1994

  20. [28]

    Fiber Orientation Estimation from X-ray Dark Field Images of Fiber Reinforced Polymers Using Constrained Spherical Deconvolution,

    B. Huyge, J. Sanctorum, B. Jeurissen, J. De Beenhouwer, and J. Sijbers, “Fiber Orientation Estimation from X-ray Dark Field Images of Fiber Reinforced Polymers Using Constrained Spherical Deconvolution,” Polymers, vol. 15, no. 13, p. 2887, Jan. 2023

    2023

  21. [29]

    Signal-to-noise ratio in x ray dark-field imaging using a grating interferometer,

    M. Chabior, T. Donath, C. David, M. Schuster, C. Schroer, and F. Pfeiffer, “Signal-to-noise ratio in x ray dark-field imaging using a grating interferometer,”Journal of Applied Physics, vol. 110, no. 5, p. 053105, Sep

  22. [30]

    Beck, First-Order Methods in Optimization

    A. Beck, First-Order Methods in Optimization. Society for Industrial and Applied Mathematics, Oct. 2017. [Online]. Available: http://dx.doi.org/10.1137/1.9781611974997

    2017 doi

  23. [31]

    Eriksson, C

    K. Eriksson, C. Johnson, and D. Estep, Applied Mathematics: Body and Soul. Springer Berlin Heidelberg, 2004. [Online]. Available: http://dx.doi.org/10.1007/978-3-662-05800-8

    2004 doi

  24. [32]

    Elsa: An elegant framework for tomographic reconstruction,

    D. Frank, J. Jelten, and T. Lasser, “Elsa: An elegant framework for tomographic reconstruction,” Journal of Open Source Software, vol. 9, no. 94, p. 6174, Feb. 2024. 12

    2024

  25. [2011]

    Available: https://doi.org/10.1063/1.3630051

    [Online]. Available: https://doi.org/10.1063/1.3630051

    doi

  26. [2024]

    Available: https://eurradiolexp.springeropen.com/articles/10.1186/s41747-023-00399-w

    [Online]. Available: https://eurradiolexp.springeropen.com/articles/10.1186/s41747-023-00399-w

    doi

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.