Pith. sign in

REVIEW 5 major objections 6 minor 47 references

Physically-Based Inverse Rendering Framework for PET Image Reconstruction

T0 review · 5 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper treats PET image reconstruction as a physically-based inverse rendering problem, using a differentiable Monte Carlo forward model and automatic differentiation to recover the emission image, and reports higher SNR than CASToR and

desk verdict A real engineering proof-of-concept for differentiable PET reconstruction, but the claims outrun the evidence: the update is just MLEM, the MC sample count is unreported, and the clinical arm is n=1. read the letter →

arxiv 2508.19956 v1 pith:BF6HRG2A submitted 2025-08-27 physics.med-ph

Yixin Li , Soroush Shabani Sichani , Zipai Wang , Wanbin Tan , Baptiste Nicolet , Xiuyuan Wang , David A. Muller , Gloria C. Chiang
show 2 more authors
Wenzel Jakob Amir H. Goldan
This is my paper
classification physics.med-ph
keywords PETimagereconstructioninverserenderingautomaticdifferentiationMonteCarlosamplingMLEMdifferentiabletime-of-flightsystemmatrix
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

The paper aims to establish that PET image reconstruction can be recast as a physically-based inverse rendering problem, where the scanner's forward process is simulated by path tracing and the emission image is optimized by automatic differentiation. The authors claim this removes the need to manually derive MLEM-style update equations and lets the system model incorporate physical effects such as crystal blurring, normalization, attenuation, and time-of-flight directly. They validate the framework on phantom and clinical brain data, reporting higher signal-to-noise ratios than the CASToR toolkit and higher hippocampal uptake ratios and gray-to-white contrast than the Siemens platform. If the claims hold, the framework offers a physically interpretable, extensible route to PET reconstruction that adapts to new scanner geometries without re-deriving analytical projectors.

What carries the argument

The central object is the differentiable forward renderer ℛ(𝐈), which maps the emission image to a sinogram by path tracing: Sobol-sampled points inside crystal volumes define sub-LORs, line integrals through the activity image are weighted by time-of-flight Gaussians, and normalization and attenuation correction maps are applied multiplicatively. Gradients of the loss with respect to the image come from automatic differentiation through this renderer, with the sensitivity image defined as the Jacobian-transpose applied to a uniform input, so the MLEM update becomes a gradient step rather than a manually derived equation.

What would settle it

Run the same list-mode phantom data through the IR framework with B set to 1 and again with B very large, and compare the reconstructed images and SNR; if quality changes substantially, the Monte Carlo projector's variance is load-bearing. A sharper test is to replace the sampled sub-LOR average with an exact analytical integration over crystal volumes and check whether the reported gains over CASToR persist, while keeping all other parameters fixed.

Watch

Extended reading notes

Core claim

The paper's central claim is that PET image reconstruction can be treated as an inverse rendering problem: the unknown radiotracer distribution is a volumetric scene parameter, and the measured sinogram is the rendered image. The forward model replaces the fixed system matrix with a hybrid Monte Carlo and analytical projector: random points sampled inside the scintillator crystals define sub-LORs, line integrals through the activity image are computed along these rays, and time-of-flight information enters as a Gaussian weight along each sub-LOR. The gradient of the Poisson log-likelihood loss with respect to every voxel is then obtained by automatic differentiation through this differentiab

Load-bearing premise

The reconstruction assumes that averaging a finite number of randomly sampled sub-LORs per detector pair gives an unbiased and sufficiently low-variance estimate of the true detector response, and the paper reports neither the sample count B nor any variance diagnostic.

Editorial extensions

If this is right

  • PET reconstruction no longer requires explicitly building or storing a system matrix; the forward model is evaluated dynamically by ray sampling, so changing scanner geometry or physics only requires modifying the renderer.
  • The same differentiable pipeline can support multiple statistical objectives—Poisson likelihood, least squares, weighted least squares, KL divergence—without re-deriving update equations for each.
  • Time-of-flight information is naturally incorporated as a Gaussian weight along each sub-LOR, making TOF resolution a direct parameter of the forward model.
  • In phantom experiments, IR reconstruction yields higher SNR across all Derenzo hot-spot sizes and lower gray-matter coefficient of variation and ventricle spill-over ratio than CASToR, indicating improved noise suppression and cold-region recovery.
  • In the clinical tau scan, IR reconstruction gives 9.97% higher hippocampal SUVR and 3.52% higher gray-to-white ratio than the Siemens Biograph mCT, with gray-matter COV higher by 3.83 percentage points.
  • Because the forward model is modular, additional physical effects such as Compton scattering, positron range, and photon acollinearity can be added as Monte Carlo components without changing the optimization machinery.

Reading between the lines

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

  • If the Monte Carlo estimator uses a small number of sub-LOR samples per detector pair, the reported SNR gains may partly reflect smoothing from the sampling itself rather than better physical modeling; reporting B and seed variance would separate the two.
  • The framework's ability to differentiate through arbitrary forward-model components suggests a natural extension to joint reconstruction of activity and attenuation maps, since attenuation enters only as a multiplicative correction in the current pipeline.
  • Because update equations are no longer hand-derived, adding a penalized prior (e.g., total variation or a learned regularizer) becomes a matter of appending a differentiable loss term; the paper mentions this possibility but does not demonstrate it.
  • The clinical comparison uses different algorithms and iteration counts (4 iterations/21 subsets OSEM versus 84 MLEM iterations), so the reported SUVR and GWR differences may not isolate the forward-model improvement from the algorithm choice.
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

5 major / 6 minor

Summary. The paper proposes a physically-based inverse rendering (IR) framework for PET image reconstruction, implemented with Dr.Jit. The forward model uses Sobol-sampled sub-LORs within detector crystal pairs, a Gaussian perturbation kernel for PSF modeling, optional TOF weighting, and normalization/attenuation correction; the inverse problem is solved by gradient-based optimization with automatic differentiation, with MLEM expressed as a gradient-ascent update. The framework is evaluated on ultra-micro Derenzo and Hoffman brain phantoms acquired on Prism-PET and on one clinical [18F]MK6240 brain scan acquired on a Siemens Biograph mCT, comparing against CASToR MLEM and the Siemens OSEM reconstruction. Reported results include higher SNR, lower COV/SOR in phantoms, mixed PVR, higher hippocampal SUVR and GWR in the single clinical case, and faster reconstruction times by 4.4-9.7x.

Significance. The contribution is a GPU-based differentiable rendering pipeline for PET that could make it easier to incorporate complex physical models without deriving gradients by hand. If the stochastic forward projector is unbiased and has controlled variance, the speedups and phantom trends are useful. However, the central quantitative claims are not yet supported: the forward-model equation contains an apparent degeneracy, the Monte Carlo sample count B is unreported, and the clinical comparison is unmatched and underpowered. No code or data are provided, and no parameter-free derivations are offered. The paper is a plausible methods demonstration but not yet a validated reconstruction platform.

major comments (5)
  1. [Section III-B, Eq. (17)] The contribution function is defined as f(r_ij(t))=W_e^i(r)Phi(r)W_e^j(r), with W_e binary indicators for crystal interaction. For t in (0,1), r_ij(t) lies between the two crystals, so both indicators vanish and L_ij in Eq. (12) is zero almost everywhere. If the indicators are meant to be endpoint weights, the equation should be written with W_e^i(p_i)W_e^j(p_j) outside the line integral; if they are meant to be pointwise, the forward model is degenerate. This must be corrected before the numerical results can be interpreted.
  2. [Section III-B, Eq. (16); all experiments] The number B of random sub-LORs per detector pair is never reported, and no bias/variance analysis or convergence check for the Monte Carlo estimator is given. The update in Eq. (22) is a MLEM iteration whose forward projection depends on this estimator; if B is small or the sub-LORs are resampled each iteration, the gradient is stochastic and the algorithm may not converge to the maximum-likelihood solution. Since all reported metrics (SNR, PVR, COV, SOR, SUVR/GWR) are computed from such reconstructions, B, the sampling/reuse policy, and a variance or convergence analysis must be provided.
  3. [Section IV-C, Table II] The clinical comparison is based on a single subject and compares a 4-iteration/21-subset OSEM reconstruction from Siemens with an 84-iteration MLEM reconstruction from the IR framework. The reported differences are within ROI standard deviations (SUVR 0.793±0.116 vs 0.872±0.131; GWR 1.254±0.225 vs 1.298±0.278), so the abstract's claim of 'higher hippocampal SUVR and GWR' is not supported. Either replace this with a matched comparison and more subjects, or explicitly label it as a single-case illustrative observation.
  4. [Section IV-A/B, Fig. 3, Table I] SNR and COV are reported as point estimates without error bars or uncertainty quantification; the SOR uncertainties are given but the source (slice variability, noise realizations, ROI definition) is not described. The differences are small (COV 8.013% vs 7.842%), and the PVR comparison is qualitatively mixed (IR higher only at the 1.00 mm diameter). Add error bars, multiple noise realizations or bootstraps, and a statement of what the error bars represent before claiming improved image quality.
  5. [Section III-B, Eq. (18); Section IV] The forward renderer in Eq. (18) includes normalization and attenuation but no additive randoms/scatter term r from Eq. (1), and the experimental protocols do not describe how randoms/scatter are handled. If the measured sinograms contain randoms/scatter, the Poisson likelihood update in Eq. (22) is mismatched. State how randoms/scatter are corrected or included, or restrict the claims to precorrected data.
minor comments (6)
  1. [Section III-B] Equation (15) is empty/missing; the numbering jumps from (14) to (16). Please supply the missing equation or renumber.
  2. [Section III-C] The symbol L is used both as a loss to minimize in Eq. (19) and as a log-likelihood to maximize in the gradient-ascent MLEM expression in Eq. (22). Clarify the sign convention or use negative log-likelihood for the loss.
  3. [Section III-B] The text calls the sub-LORs 'randomly sampled' but uses Sobol quasi-Monte Carlo sampling. Clarify whether the samples are deterministic low-discrepancy sequences or stochastic, as this affects the variance discussion.
  4. [Table III] PSF FWHM values are given without units, and the number of iterations for the Hoffman phantom (80) is not stated in Section IV-B. Please harmonize.
  5. [Section V.D and Fig. 6] The speedup is quoted as '4.5x' in the text and '4.4x' in the figure caption for the Hoffman phantom. Please make the numbers consistent.
  6. [Abstract] The phrase 'first ever platform' is a strong claim; suggest softening to 'to our knowledge, the first...' unless a comprehensive prior-art search is provided.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the MLEM update is explicit, the forward model is independently defined, and all quantitative claims are empirical benchmarks rather than outputs of a fitted/self-referential chain.

full rationale

The derivation chain is self-contained. The forward model is defined in Eqs. (8)-(18) as a line-integral/Monte Carlo renderer of the activity image, with normalization/attenuation corrections; it is not defined in terms of the target metrics or the measured sinogram in a way that would force the reported improvements. The inverse step uses the standard MLEM update (Eq. 22, citing Kaufman [39]), with the sensitivity image defined as the gradient of the forward model with respect to a uniform input (Eq. 23). This is algebraically the standard MLEM equation (Eq. 4), not a circular reformulation: the gradient is computed by AD from the same forward model, and no parameter in the chain is fitted to the SNR/SUVR/GWR endpoints. The comparisons to CASToR and Siemens Biograph mCT are empirical evaluations on phantom and clinical data; the improved metrics are measured outcomes of a particular implementation, not consequences of the definitions. Self-citations [33][34] are prior-work references used as background for differentiable rendering in PET and are not load-bearing; [35] (Dr.Jit) is a tool citation. The unreported sub-LOR sample count B in Eq. (16) is a potential robustness/validity gap in the Monte Carlo estimator, but it does not make any claimed result equivalent to its inputs by construction, so it does not constitute circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The framework introduces no new physical entities. The free parameters are hand-chosen reconstruction settings and an unstated Monte Carlo sample count. The axioms are standard PET modeling assumptions plus a few domain-specific choices about forward model adequacy and comparison fairness. The unstated sample count B is the most concerning hidden parameter because it directly affects the noise and bias of the projection operator.

free parameters (3)
  • PSF FWHM (Gaussian perturbation kernel) = 1.2 mm (Derenzo), 1.5 mm (Hoffman)
    Set by hand to model spatial resolution; chosen per experiment, not fitted to data.
  • Number of Monte Carlo sub-LOR samples B
    Not reported in the paper; a critical parameter controlling estimator variance and bias of the forward model.
  • Number of MLEM iterations = 200 (Derenzo), 80 (Hoffman), 84 (clinical)
    Chosen by hand; no convergence criterion or stopping rule is given.
assumptions (5)
  • standard math PET data are independent Poisson random variables (Eq 2)
    Standard assumption in PET reconstruction; stated in Section II-A.
  • domain assumption The forward model (line integrals with normalization and attenuation multiplication) is an adequate representation of the PET system response
    Assumed throughout; not validated against a measured system matrix, Section III-B.
  • standard math The MLEM update in Eq 22 is equivalent to the standard MLEM algorithm and converges to a maximum likelihood estimate
    Relies on Kaufman [39] for the gradient ascent formulation; no convergence proof is given in the paper.
  • domain assumption The Monte Carlo estimator (Eq 16) with B subsamples provides an unbiased estimate of the line integral response
    Depends on B and sampling strategy; B is not reported, so unbiasedness and low variance are not established.
  • domain assumption The clinical SUVR/GWR comparison assumes the Siemens Biograph mCT OSEM reconstruction is a fair baseline for tissue contrast
    The two reconstructions use different algorithms, iteration counts, and possibly different post-processing, Section V-C.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physically-Based Inverse Rendering Framework for PET Image Reconstruction." pith.science (2026). https://pith.science/paper/BF6HRG2A

@misc{pith2026250819956,
  author       = {Pith},
  title        = {Pith review of: Physically-Based Inverse Rendering Framework for PET Image Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BF6HRG2A}},
  note         = {Machine review of arXiv:2508.19956}
}
read the original abstract

Differentiable rendering has been widely adopted in computer graphics as a powerful approach to inverse problems, enabling efficient gradient-based optimization by differentiating the image formation process with respect to millions of scene parameters. Inspired by this paradigm, we propose a physically-based inverse rendering (IR) framework, the first ever platform for PET image reconstruction using Dr.Jit, for PET image reconstruction. Our method integrates Monte Carlo sampling with an analytical projector in the forward rendering process to accurately model photon transport and physical process in the PET system. The emission image is iteratively optimized using voxel-wise gradients obtained via automatic differentiation, eliminating the need for manually derived update equations. The proposed framework was evaluated using both phantom studies and clinical brain PET data acquired from a Siemens Biograph mCT scanner. Implementing the Maximum Likelihood Expectation Maximization (MLEM) algorithm across both the CASToR toolkit and our IR framework, the IR reconstruction achieved a higher signal-to-noise ratio (SNR) and improved image quality compared to CASToR reconstructions. In clinical evaluation compared with the Siemens Biograph mCT platform, the IR reconstruction yielded higher hippocampal standardized uptake value ratios (SUVR) and gray-to-white matter ratios (GWR), indicating enhanced tissue contrast and the potential for more accurate tau localization and Braak staging in Alzheimer's disease assessment. The proposed IR framework offers a physically interpretable and extensible platform for high-fidelity PET image reconstruction, demonstrating strong performance in both phantom and real-world scenarios.

Figures

Figures reproduced from arXiv: 2508.19956 by the authors.

Figure 2
Figure 2. Reconstructed images of ultra-micro Derenzo phantoms using the CASToR reconstruction toolkit and the proposed IR framework [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Representative axial slices of the reconstructed Hoffman brain [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 6
Figure 6. Mean run time for Hoffman Brain and Hotspot phantoms. Inverse Rendering outperforms CASToR by 4.4x and 9.7x on Hoffman and Hotspot phantoms respectively. TABLE III Reconstruction Parameters used in CASToR and IR Hoffman Hotspot PSF FWHM 1.5 1.2 Number of iterations 80 200 VI. DISCUSSION The IR reconstruction demonstrated superior noise suppression, as evidenced by a significantly lower COV in the gray matter of the … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 45 canonical work pages

  1. [1]

    Positron emission tomography provides molecular imaging of biological processes,

    M. E. Phelps, “Positron emission tomography provides molecular imaging of biological processes,” Proc. Natl. Acad. Sci. U.S.A., vol. 97, no. 16, pp. 9226–9233, Aug. 2000

    work page 2000

  2. [2]

    R. A. Robb, Ed., Three -Dimensional Biomedical Imaging (1985), Volume II, 1st ed. Boca Raton, FL, USA: CRC Press, 1985

    work page 1985

  3. [3]

    Analytic 3D image reconstruction using all detected events,

    P. E. Kinahan and J. G. Rogers, “Analytic 3D image reconstruction using all detected events,” IEEE Trans. Nucl. Sci., vol. 36, no. 1, pp. 964–968, Feb. 1989

    work page 1989

  4. [4]

    Theoretically exact filtered backprojection-type inversion algorithm for spiral CT,

    A. Katsevich, “Theoretically exact filtered backprojection-type inversion algorithm for spiral CT,” SIAM J. Appl. Math., vol. 62, no. 6, pp. 2012– 2026, Dec. 2002

    work page 2012

  5. [5]

    Fully-three-dimensional positron emission tomography,

    J. G. Colsher, “Fully-three-dimensional positron emission tomography,” Phys. Med. Biol., vol. 25, no. 1, pp. 103–115, Jan. 1980

    work page 1980

  6. [6]

    Evaluation of 3D reconstruction algorithms for a small animal PET camera,

    C. A. Johnson et al., “Evaluation of 3D reconstruction algorithms for a small animal PET camera,” IEEE Trans. Nucl. Sci., vol. 44, no. 3, pp. 1303–1308, Jun. 1997

    work page 1997

  7. [7]

    Noise correlation in PET, CT, SPECT and PET/CT data evaluated using autocorrelation function: A phantom study on data, reconstructed using FBP and OSEM,

    P. Razifar et al., “Noise correlation in PET, CT, SPECT and PET/CT data evaluated using autocorrelation function: A phantom study on data, reconstructed using FBP and OSEM,” BMC Med. Imaging, vol. 5, no. 1, Art. no. 5, Aug. 2005

    work page 2005

  8. [8]

    Performance evaluation of filtered backprojection reconstruction and iterative reconstruction methods for PET images,

    C. X. Wang, W. E. Snyder, G. Bilbro, and P. Santago, “Performance evaluation of filtered backprojection reconstruction and iterative reconstruction methods for PET images,” Comput. Biol. Med., vol. 28, no. 1, pp. 13–25, Jan. 1998

    work page 1998

Show all 47 references
  1. [9]

    Iterative reconstruction techniques in emission computed tomography,

    J. Qi and R. M. Leahy, “Iterative reconstruction techniques in emission computed tomography,” Phys. Med. Biol., vol. 51, no. 15, pp. R541 – R578, Aug. 2006

    2006

  2. [10]

    High-resolution 3D Bayesian image reconstruction using the microPET small-animal scanner,

    J. Qi, R. M. Leahy, S. R. Cherry, A. Chatziioannou, and T. H. Farquhar, “High-resolution 3D Bayesian image reconstruction using the microPET small-animal scanner,” Phys. Med. Biol., vol. 43, no. 4, pp. 1001 –1013, Apr. 1998

    1998

  3. [11]

    Advances in PET image reconstruction,

    A. J. Reader and H. Zaidi, “Advances in PET image reconstruction,” PET Clin., vol. 2, no. 2, pp. 173–190, Apr. 2007

    2007

  4. [12]

    200x low -dose PET reconstruction using deep learning,

    J. Xu, E. Gong, J. M. Pauly, and G. Zaharchuk, “200x low -dose PET reconstruction using deep learning,” arXiv:1712.04119, Dec. 2017. [Online]. Available: http://arxiv.org/abs/1712.04119

    2017 arXiv

  5. [13]

    Artificial intelligence and deep learning for advancing PET image reconstruction: State -of-the-art and future directions,

    D. Hellwig, N. C. Hellwig, S. Boehner, T. Fuchs, R. Fischer, and D. Schmidt, “Artificial intelligence and deep learning for advancing PET image reconstruction: State -of-the-art and future directions,” Nuklearmedizin, vol. 62, no. 6, pp. 334–342, Dec. 2023

    2023

  6. [14]

    DeepPET: A deep encoder–decoder network for directly solving the PET image reconstruction inverse problem,

    I. Häggström, C. R. Schmidtlein, G. Campanella, and T. J. Fuchs, “DeepPET: A deep encoder–decoder network for directly solving the PET image reconstruction inverse problem,” Med. Image Anal., vol. 54, pp. 253–262, May 2019

    2019

  7. [15]

    Deep learning based framework for direct reconstruction of PET images,

    Z. Liu, H. Chen, and H. Liu, “Deep learning based framework for direct reconstruction of PET images,” in Proc. MICCAI, Shenzhen, China, Oct. 2019, vol. 11766, Lect. Notes Comput. Sci., pp. 48–56

    2019

  8. [16]

    DirectPET: Full -size neural network PET reconstruction from sinogram data,

    W. Whiteley, W. K. Luk, and J. Gregor, “DirectPET: Full -size neural network PET reconstruction from sinogram data,” J. Med. Imaging, vol. 7, no. 3, 032503, Feb. 2020

    2020

  9. [17]

    PET image reconstruction incorporating deep image prior and a forward projection model,

    F. Hashimoto, K. Ote, and Y. Onishi, “PET image reconstruction incorporating deep image prior and a forward projection model,” IEEE Trans. Radiat. Plasma Med. Sci., vol. 6, no. 8, pp. 841–846, Nov. 2022

    2022

  10. [18]

    Fully 3D implementation of the end -to-end deep image prior -based PET image reconstruction using block iterative algorithm,

    F. Hashimoto, Y. Onishi, K. Ote, H. Tashima, and T. Yamaya, “Fully 3D implementation of the end -to-end deep image prior -based PET image reconstruction using block iterative algorithm,” Phys. Med. Biol., vol. 68, no. 15, 155009, Jul. 2023

    2023

  11. [19]

    Iterative PET image reconstruction using convolutional neural network representation,

    K. Gong et al., “Iterative PET image reconstruction using convolutional neural network representation,” IEEE Trans. Med. Imaging, vol. 38, no. 3, pp. 675–685, Mar. 2019

    2019

  12. [20]

    Deep learning applications for quantitative and qualitative PET in PET/MR: technical and clinical unmet needs,

    J. Yang, A. Afaq, R. Sibley, A. McMilan, and A. Pirasteh, “Deep learning applications for quantitative and qualitative PET in PET/MR: technical and clinical unmet needs,” Magn. Reson. Mater. Phys., Biol. Med., vol. 37, no. 4, pp. 749–763, Aug. 2024

    2024

  13. [21]

    A review on AI in PET imaging,

    K. Matsubara, M. Ibaraki, M. Nemoto, H. Watabe, and Y. Kimura, “A review on AI in PET imaging,” Ann. Nucl. Med., vol. 36, no. 2, pp. 133– 143, Feb. 2022

    2022

  14. [22]

    Differentiable rendering: A survey,

    H. Kato, D. Beker, M. Morariu, T. Ando, T. Matsuoka, W. Kehl, and A. Gaidon, “Differentiable rendering: A survey,” arXiv:2006.12057, 2020. [Online]. Available: https://arxiv.org/abs/2006.12057

    2006 arXiv

  15. [23]

    Differentiable Monte Carlo ray tracing through edge sampling,

    T.-M. Li, M. Aittala, F. Durand, and J. Lehtinen, “Differentiable Monte Carlo ray tracing through edge sampling,” ACM Trans. Graph., vol. 37, no. 6, Art. no. 222, Dec. 2018

    2018

  16. [24]

    Extended path space manifolds for physically based differentiable rendering,

    J. Xing, X. Hu, F. Luan, L. -Q. Yan, and K. Xu, "Extended path space manifolds for physically based differentiable rendering," SIGGRAPH Asia 2023 Conference Papers, New York, NY, USA: Association for Computing Machinery, Art. no. 30, 2023

    2023

  17. [25]

    A survey of inverse rendering problems,

    G. Patow and X. Pueyo, “A survey of inverse rendering problems,” Comput. Graph. Forum, vol. 22, no. 4, pp. 663–687, 2003

    2003

  18. [26]

    Large steps in inverse rendering of geometry,

    B. Nicolet, A. Jacobson, and W. Jakob, “Large steps in inverse rendering of geometry,” ACM Trans. Graph., vol. 40, no. 6, Art. no. 248, Dec. 2021

    2021

  19. [27]

    Physics-based differentiable rendering: From theory to implementation,

    S. Zhao, W. Jakob, and T.-M. Li, “Physics-based differentiable rendering: From theory to implementation,” in ACM SIGGRAPH 2020 Courses, Virtual Event, USA: Association for Computing Machinery, 2020, Art. no. 14

    2020

  20. [28]

    Pharr, W

    M. Pharr, W. Jakob, and G. Humphreys, Physically Based Rendering: From Theory to Implementation. Cambridge, MA: MIT Press, 2023

    2023

  21. [29]

    Solving inverse PDE problems using grid-free Monte Carlo estimators,

    E. F. Yilmazer, D. Vicini, and W. Jakob, “Solving inverse PDE problems using grid-free Monte Carlo estimators,” ACM Trans. Graph., vol. 43, no. 6, Art. no. 175, Dec. 2024

    2024

  22. [30]

    Image reconstruction for PET/CT scanners: past achievements and future challenges,

    S. Tong, A. M. Alessio, and P. E. Kinahan, “Image reconstruction for PET/CT scanners: past achievements and future challenges,” Imaging Med., vol. 2, no. 5, pp. 529–545, Oct. 2010

    2010

  23. [31]

    EM reconstruction algorithms for emission and transmission tomography,

    K. Lange and R. Carson, “EM reconstruction algorithms for emission and transmission tomography,” J. Comput. Assist. Tomogr., vol. 8, no. 2, pp. 306–316, Apr. 1984

    1984

  24. [32]

    Robust Monte Carlo methods for light transport simulation,

    E. Veach, “Robust Monte Carlo methods for light transport simulation,” Ph.D. dissertation, Dept. Comput. Sci., Stanford Univ., Stanford, CA, 1998

    1998

  25. [33]

    Inverse rendering-based method for PET image reconstruction,

    Y. Li, Z. Wang, W. Jakob, W. Tan, B. Nicolet, and A. Goldan, “Inverse rendering-based method for PET image reconstruction,” J. Nucl. Med., vol. 66, no. suppl. 1, p. 252047, June 2025. [Online]. Available: https://jnm.snmjournals.org/content/66/supplement_1/252047

    2025

  26. [34]

    Inverse rendering for PET image reconstruction,

    W. Jakob, Y. Li, Z. Wang, B. Nicolet, W. Tan, and A. Goldan, “Inverse rendering for PET image reconstruction,” in Proc. IEEE Nucl. Sci. Symp. Med. Imag. Conf. Room Temperature Semiconductor Detector Conf. (NSS/MIC/RTSD), Tampa, FL, USA, 2024

    2024

  27. [35]

    Dr.Jit: A just -in-time compiler for differentiable rendering,

    W. Jakob, S. Speierer, N. Roussel, and D. Vicini, “Dr.Jit: A just -in-time compiler for differentiable rendering,” ACM Trans. Graph., vol. 41, no. 4, Art. no. 124, Jul. 2022

    2022

  28. [36]

    Quasi -Monte Carlo integration,

    W. J. Morokoff and R. E. Caflisch, “Quasi -Monte Carlo integration,” J. Comput. Phys., vol. 122, no. 2, pp. 218–230, 1995

    1995

  29. [37]

    Update on time-of-flight PET imaging,

    S. Surti, “Update on time-of-flight PET imaging,” J. Nucl. Med., vol. 56, no. 1, pp. 98–105, Jan. 2015

    2015

  30. [38]

    Recent developments in time-of-flight PET,

    S. Vandenberghe, E. Mikhaylova, E. D’Hoe, P. Mollet, and J. S. Karp, “Recent developments in time-of-flight PET,” EJNMMI Phys., vol. 3, no. 1, Art. no. 3, Dec. 2016

    2016

  31. [39]

    Maximum likelihood, least squares, and penalized least squares for PET,

    L. Kaufman, “Maximum likelihood, least squares, and penalized least squares for PET,” IEEE Trans. Med. Imaging, vol. 12, no. 2, pp. 200–214, Jun. 1993

    1993

  32. [40]

    A conformal TOF–DOI Prism-PET prototype scanner for high-resolution quantitative neuroimaging,

    X. Zeng et al., “A conformal TOF–DOI Prism-PET prototype scanner for high-resolution quantitative neuroimaging,” Med. Phys., vol. 50, no. 6, pp. 3401–3417, Jun. 2023

    2023

  33. [41]

    Depth- encoding using optical photon TOF in a prism-PET detector with tapered crystals,

    X. Zeng, A. LaBella, Z. Wang, Y. Li, W. Tan, and A. H. Goldan, “Depth- encoding using optical photon TOF in a prism-PET detector with tapered crystals,” Med. Phys., vol. 51, no. 6, pp. 4044–4055, Jun. 2024

    2024

  34. [42]

    High-resolution depth-encoding PET detector module with prismatoid light -guide array,

    A. LaBella et al., “High-resolution depth-encoding PET detector module with prismatoid light -guide array,” J. Nucl. Med., vol. 61, no. 10, pp. 1528–1533, Oct. 2020

    2020

  35. [43]

    CASToR: A generic data organization and processing code framework for multi -modal and multi -dimensional tomographic reconstruction,

    T. Merlin et al., “CASToR: A generic data organization and processing code framework for multi -modal and multi -dimensional tomographic reconstruction,” Phys. Med. Biol., vol. 63, no. 18, p. 185005, Sept. 2018

    2018

  36. [44]

    System models for PET statistical iterative reconstruction: A review,

    A. Iriarte, R. Marabini, S. Matej, C. O. S. Sorzano, and R. M. Lewitt, “System models for PET statistical iterative reconstruction: A review,” Comput. Med. Imaging Graph., vol. 48, pp. 30–48, Mar. 2016

    2016

  37. [45]

    Accurate geometric and physical response modelling for statistical image reconstruction in high resolution PET,

    E. U. Mumcuoglu, R. M. Leahy, S. R. Cherry, and E. Hoffman, “Accurate geometric and physical response modelling for statistical image reconstruction in high resolution PET,” in Proc. IEEE Nucl. Sci. Symp. Conf. Rec., Anaheim, CA, USA, 1996, vol. 3, pp. 1569–1573

    1996

  38. [46]

    (2025, Feb.)

    NVIDIA Corp. (2025, Feb.). Parallel Thread Execution ISA, Version 9. [Online]. Available: https://docs.nvidia.com/cuda/parallel -thread- execution

    2025

  39. [47]

    Self -supervised and supervised deep learning for PET image reconstruction,

    A. J. Reader, “Self -supervised and supervised deep learning for PET image reconstruction,” AIP Conf. Proc., vol. 3061, no. 1, Mar. 2024, Art. no. 030003

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.