Pith. sign in

REVIEW 4 major objections 4 minor 36 references

Ultrafast Deep Learning-Based Scatter Estimation in Cone-Beam Computed Tomography

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read A 40×32 projection input makes deep-learning CBCT scatter estimation 78× cheaper and 16× faster with slightly better accuracy.

desk verdict Useful efficiency result for CBCT scatter correction, but the accuracy claim is only established in simulation and model selection is done on the test set; worth a serious referee with fixes. read the letter →

arxiv 2509.08973 v1 pith:CPFENYYE submitted 2025-09-10 eess.SP cs.CV

classification eess.SPcs.CV
keywords cone-beamCTscatterestimationdeeplearningU-NetdownsamplingbicubicinterpolationMonteCarlosimulationedgedevices
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 claims that because X-ray scatter in CBCT is a low-frequency signal, feeding a U-Net scatter estimator with heavily downsampled projections—40×32 pixels instead of 320×256—cuts floating-point operations by 78×, inference time by 16×, and GPU memory by 12× while slightly improving scatter-estimation accuracy: MAPE 3.85% vs 4.42% and MSE 1.34e-2 vs 2.01e-2. The authors establish this by training the same architecture at five resolutions on Monte Carlo simulated scatter–primary pairs and validating on large simulated datasets and real phantom scans. A good-faith reader would take the central claim to be that downsampling is an underused, effectively free lever for making deep-learning scatter correction practical on mobile and edge CBCT devices.

What carries the argument

The machinery is a U-Net scatter estimator (Aux-Net) whose encoder is given auxiliary channels encoding normalized field-of-measurement width and height, fed with linearized projections downsampled by bicubic interpolation to a fixed low resolution; the output scatter is upsampled back to the original projection size and subtracted from the projection. The argument rests on treating scatter as low-frequency: bicubic interpolation (16-neighbor cubic convolution) proved the most faithful down/up resampler among the four tested, and the network's resolution sweep reveals that 40×32 sits before accuracy falls off.

What would settle it

Measure the scatter on a real CBCT scanner with a beam-blocker (or otherwise direct) method at the same FOM settings used in simulation, and compare the 40×32 network's predicted scatter to the measured scatter; if the real-scatter MAPE is much worse than the simulated 3.85% or no better than the 320×256 baseline, the transfer premise fails.

Watch

Extended reading notes

Core claim

The central discovery is an optimum: at a 40×32 input, the network (four downsampling blocks, 1.8M parameters) achieves the lowest projection-domain MSE and matches the baseline MAPE, while at 20×16 accuracy degrades. The paper shows this holds across 30 field-of-measurement sizes in simulation and on real water and SedentexCT phantoms, where the corrected images reduce cupping; it also reports that a small water bottle is overcorrected (uniformity worsens from 8 HU to about 16 HU), a limitation the authors attribute to the training data lacking such small objects.

Load-bearing premise

The entire supervised pipeline is trained on Monte Carlo simulated scatter–primary pairs, and the real-scanner evidence is qualitative; the claim that the simulated accuracy (3.85% MAPE) transfers to physical CBCT scans is load-bearing, and the paper's own small-water-bottle result (uniformity worsening from 8 to 16–17 HU) shows the transfer is not perfect.

Editorial extensions

If this is right

  • On simulated test data, net-40×32 achieves lower MSE (1.34e-2 vs 2.01e-2) and comparable MAPE to the 320×256 baseline, so the speed gain is not bought with accuracy.
  • Reconstruction RMSE averaged over 30 FOM sizes is 8.96 HU for net-40×32 versus 9.66 for the baseline, so scatter-corrected image quality is preserved or slightly improved.
  • Reducing input to 40×32 also lets the network shrink from 7.3M to 1.8M parameters, cutting GPU memory from 3.9 GB to 310 MB and enabling batch processing on commodity cards.
  • Bicubic interpolation should replace nearest-neighbor, area, and bilinear choices for the down-up sampling step in future scatter-estimation pipelines.

Reading between the lines

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

  • Because the scatter field is low-frequency, the same downsampling idea may push further along the angular axis—training on fewer projections per scan—although the paper only mentions this as future work.
  • The 40×32 optimum may be architecture-dependent; the mechanism (low-frequency prior) suggests other U-Net variants would show a similar plateau, but the exact sweet spot should be re-measured.
  • The open transfer risk is the simulator: a beam-stop phantom measurement on a real CBCT unit would let someone quantify how much of the simulated accuracy survives in the clinic, beyond the qualitative phantom scans shown.
  • The small-water-bottle overcorrection hints that training distribution should include small, low-scatter objects; if real deployments include such cases, the model may need augmentation rather than a resolution change.
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 that deep-learning-based CBCT scatter estimation networks can be run on severely downsampled projections without loss of accuracy, enabling deployment on resource-limited hardware. The authors (1) compare four interpolation methods for down/upsampling the scatter signal at six resolutions, (2) train a recent U-Net architecture (Aux-Net) at five input resolutions with corresponding reductions in network depth/parameters, and (3) evaluate the resulting models on large simulated MC datasets as well as on real water and SedentexCT phantom scans. The central quantitative claim is that the net-40×32 configuration reduces FLOPs by 78x, inference time by 16x, and GPU memory by 12x compared with the 320×256 baseline, while slightly improving MAPE (3.85% vs. 4.42%) and MSE (1.34e-2 vs. 2.01e-2).

Significance. If the claims hold, the work offers a simple and practical lever—input downsampling with bicubic interpolation—for making deep-learning scatter correction feasible on mobile CBCT and edge devices. The study is extensive in simulation: 270,000 training projections across 18 FOM sizes, 90,000 test projections across 30 FOM sizes, 5-fold cross-validation, and resource metrics (FLOPs, inference time, memory) are explicitly reported. The use of public CT data (HNSCC-3DCT-RT) and real phantom scans is a strength. However, the quantitative claims are currently clouded by internal table inconsistencies, the optimal resolution is selected on the same simulated test set used for reporting, and real-scanner validation is only qualitative and even shows a counter-example (small water bottle overcorrection). These issues are fixable but require revision and, ideally, a quantitative real-scatter experiment.

major comments (4)
  1. [3.2, Table 2, Section 4] The quantitative basis of the headline comparison is internally inconsistent. Table 2 lists MSE under the heading (10^-2) as 2.01±0.14, 1.56±0.13, 1.43±0.08, 1.34±0.09, etc., but Section 3.2 reports net-160×128 as “MSE of 1.56±0.13×10^-1” (ten times larger) and Section 4 reports net-40×32 as “0.134±0.009” and net-320×256 as “0.20±0.014” (i.e., 1.34e-2 and 2.0e-2). Moreover, Table 2 contains two rows both labeled net-20×16, one with 0.5M parameters/0.01 GFLOPs and one with 1.8M parameters/0.04 GFLOPs; the latter appears to be a different configuration, possibly net-40×32. These must be corrected and the unit convention unified before the 78x/16x/12x claim can be checked.
  2. [2.5, 3.3] The optimal resolution is selected on the same simulated test set used to report the final errors. Section 2.5 describes only the training/validation split inside the 5-fold cross-validation; Section 3.3 then compares all five trained models on the test set and declares net-40×32 optimal. This makes the reported test errors act as model-selection criteria, so the “optimal” claim is not out-of-sample. Please either use a separate holdout set for selection/reporting or present the performance on the validation folds as the basis for the resolution choice.
  3. [3.5, Fig. 8d, 2.1] The accuracy-parity claim is demonstrated only in the MC-simulated domain. For real Viso G7 scans there is no scatter ground truth, and Fig. 8d shows that for the small water bottle all corrected networks worsen uniformity relative to uncorrected (8 HU to 14–27 HU), consistent with overcorrection. In addition, the simulations explicitly exclude bow-tie and grid (Section 2.1), while the target device may include such components. Because the central practical claim is that the low-resolution network can replace the baseline on real CBCT, a quantitative real-scatter validation (e.g., beam-stop or edge-blocker measurements, or a validated MC pipeline including the missing hardware) is needed to support the transferability of the reported gains.
  4. [Table 3] Table 3 lists the FOM label 150×100 twice, with different RMSE values, and omits 150×110 that appears in the test-set list (Table A2). Since this table supports the reconstruction-domain claim that net-40×32 is “second lowest RMSE,” the duplicate/omission must be corrected and the mean/standard deviation recomputed over the true 30 FOM sizes.
minor comments (4)
  1. [3.3] “net-320×160” should be “net-320×256”.
  2. [Abstract] Typo: “comarable” should be “comparable.”
  3. [3.1] The sentence “the scatter reconstruction error increases sharply for a factor of 2 and a factor of 32” is ambiguous; the intended comparison between factor 2 and larger factors should be rephrased for clarity.
  4. [Fig. 7, Section 4] The comparison in Fig. 7 mixes test-time MSE on the network models with interpolation-only MSE on the training data. Please clarify the axes, the data split, and the normalization so the reader can interpret the claimed divergence between interpolation error and network test error.

Circularity Check

0 steps flagged · score 2.0 of 10

Mild self-citation in architecture and MC pipeline, but the central downsampling claim is measured on held-out data, not derived from its inputs.

full rationale

The paper's derivation chain is: Monte Carlo simulation of primary/scatter pairs (Section 2.1) -> supervised training of the authors' own Aux-Net at five resolutions (Sections 2.4-2.5) -> evaluation on 30 FOM sizes excluded from training and 6 unseen scans (Section 3.3-3.4) -> real phantom scans (Section 3.5). The headline numbers (MAPE 3.85% vs 4.42%, MSE 1.34e-2 vs 2.01e-2) are direct measurements of independently trained networks on a held-out simulated test set, so they cannot reduce to the training inputs by construction. The baseline net-320x256 is retrained and re-measured in this paper (Table 2), so the comparison to 'state-of-the-art' is internally grounded rather than imported from the self-cited Aux-Net paper [20]. Self-citation exists (architecture from [20], MC pipeline following [28], SPIE paper [22]), but the load-bearing argument -- that resolution can be reduced without losing accuracy -- is empirically verified here, and the paper openly reports a counter-signal (water-bottle uniformity worsening from 8 HU to 16-17 HU, Fig. 8d), which is inconsistent with a rigged or self-confirming evaluation. The divergence between interpolation-only MSE and network test MSE (Fig. 7) shows the finding is not a renamed restatement of interpolation error. The strongest vulnerability of the paper, sim-to-real transfer of the MC-trained models, is a correctness risk about external validity, not a circularity of the derivation itself.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central result is an empirical architecture and resolution choice. It rests on a smoothness assumption about scatter, on transfer from simulation to real scanning, and on the validity of the underlying U-Net; these are domain assumptions rather than mathematical axioms. The selected input resolution and per-resolution network depth are hand-chosen using test data, so they are recorded as free parameters. No new physical entities are introduced.

free parameters (3)
  • Input resolution for scatter network = 40x32 (selected from 320x256, 160x128, 80x64, 40x32, 20x16)
    Optimal resolution chosen by comparing all five trained models on the same simulated test set (Table 2), without a separate held-out set for this choice.
  • Network depth (number of downsampling blocks) = 5 blocks for 320x256/160x128/80x64, 4 for 40x32, 3 for 20x16; Table 2 also shows an unexplained 4-block 20x16
    Depth was reduced by hand to match smaller feature maps and affects the FLOPs versus accuracy trade-off reported.
  • Interpolation method for resizing = bicubic
    Chosen as best among nearest, area, bilinear and bicubic by interpolation MSE on training scatter signals (Section 3.1).
assumptions (4)
  • domain assumption Scatter signal is sufficiently low-frequency that information needed for correction survives downsampling to 40x32.
    Stated as motivation in the Introduction and Discussion; Fig. 2 shows interpolation error grows immediately with downsampling, so the assumption is only partially validated by the network results.
  • domain assumption Monte Carlo simulated scatter-primary pairs are a valid proxy for physical CBCT scatter on Viso G7.
    All training targets come from PENELOPE-style simulations (Section 2.1); real phantom tests are qualitative and show overcorrection in a small bottle, so transfer is not fully established.
  • domain assumption Aux-Net architecture from prior work by the same authors is an appropriate state-of-the-art backbone.
    Borrowed from [20]; the paper's resource comparison inherits the architecture's design choices.
  • domain assumption FDK reconstruction and HU uniformity or RMSE capture clinically relevant image quality.
    Evaluation metrics in Sections 2.5 and 3.4 may not reflect diagnostic performance; for example, water bottle uniformity worsens after correction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ultrafast Deep Learning-Based Scatter Estimation in Cone-Beam Computed Tomography." pith.science (2026). https://pith.science/paper/CPFENYYE

@misc{pith2026250908973,
  author       = {Pith},
  title        = {Pith review of: Ultrafast Deep Learning-Based Scatter Estimation in Cone-Beam Computed Tomography},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CPFENYYE}},
  note         = {Machine review of arXiv:2509.08973}
}
read the original abstract

Purpose: Scatter artifacts drastically degrade the image quality of cone-beam computed tomography (CBCT) scans. Although deep learning-based methods show promise in estimating scatter from CBCT measurements, their deployment in mobile CBCT systems or edge devices is still limited due to the large memory footprint of the networks. This study addresses the issue by applying networks at varying resolutions and suggesting an optimal one, based on speed and accuracy. Methods: First, the reconstruction error in down-up sampling of CBCT scatter signal was examined at six resolutions by comparing four interpolation methods. Next, a recent state-of-the-art method was trained across five image resolutions and evaluated for the reductions in floating-point operations (FLOPs), inference times, and GPU memory requirements. Results: Reducing the input size and network parameters achieved a 78-fold reduction in FLOPs compared to the baseline method, while maintaining comarable performance in terms of mean-absolute-percentage-error (MAPE) and mean-square-error (MSE). Specifically, the MAPE decreased to 3.85% compared to 4.42%, and the MSE decreased to 1.34 \times 10^{-2} compared to 2.01 \times 10^{-2}. Inference time and GPU memory usage were reduced by factors of 16 and 12, respectively. Further experiments comparing scatter-corrected reconstructions on a large, simulated dataset and real CBCT scans from water and Sedentex CT phantoms clearly demonstrated the robustness of our method. Conclusion: This study highlights the underappreciated role of downsampling in deep learning-based scatter estimation. The substantial reduction in FLOPs and GPU memory requirements achieved by our method enables scatter correction in resource-constrained environments, such as mobile CBCT and edge devices.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 1 linked inside Pith

  1. [1]

    Faccioli N, Santi E, Foti G, Mansueto G, and Corain M. Cost-effectiveness of introducing cone-beam computed tomography (CBCT) in the management of complex phalangeal fractures: economic simulation.Musculoskelet Surg., 106:169– 177, 2022

  2. [2]

    Dental cone beam CT and its justified use in oral health care.J Belg Soc Radiol., 94(5):254–265, 2011

    Jacobs R. Dental cone beam CT and its justified use in oral health care.J Belg Soc Radiol., 94(5):254–265, 2011

  3. [3]

    Introduction of portable computed tomography scanners, in the treatment of acute stroke patients via telemedicine in remote communities.Int J Stroke, 5(2):62–66, 2010

    Shuaib A, Khan K, Whittaker T, Amlani S, and Crumley P. Introduction of portable computed tomography scanners, in the treatment of acute stroke patients via telemedicine in remote communities.Int J Stroke, 5(2):62–66, 2010

  4. [4]

    Cone-beam computed tomography with a flat- panel imager: magnitude and effects of x-ray scatter.Med Phys., 28(2):220–231, 2001

    Siewerdsen JH and Jaffray DA. Cone-beam computed tomography with a flat- panel imager: magnitude and effects of x-ray scatter.Med Phys., 28(2):220–231, 2001

  5. [5]

    The effects of scatter in x-ray computed tomography

    Joseph PM and Spital RD. The effects of scatter in x-ray computed tomography. Med Phys., 9(4):464–472, 1982

  6. [6]

    Kalender W A and Kyriakou. Y. Flat-detector computed tomography (FD-CT). Eur Radiol., 17:2767–2779, 2007

  7. [7]

    Scatter rejection by air gaps: An empirical model.Med Phys., 12(3):308–316, 1985

    Sorenson JA and Floch J. Scatter rejection by air gaps: An empirical model.Med Phys., 12(3):308–316, 1985

  8. [8]

    Efficiency of antiscatter grids for flat-detector CT

    Kyriakou Y and Kalender W. Efficiency of antiscatter grids for flat-detector CT. Phys Med Biol., 52(20):6275, 2007

Show all 36 references
  1. [9]

    Scatter correction for cone-beam CT in radiation therapy.Med Phys., 36(6Part1):2258–2268, 2009

    Zhu L, Xie Y, Wang J, and Xing L. Scatter correction for cone-beam CT in radiation therapy.Med Phys., 36(6Part1):2258–2268, 2009

  2. [10]

    The effects of compensator and imaging geometry on the distribution of x-ray scatter in CBCT.Med Phys., 38(2):897–914, 2011

    Bootsma GJ, Verhaegen F, and Jaffray DA. The effects of compensator and imaging geometry on the distribution of x-ray scatter in CBCT.Med Phys., 38(2):897–914, 2011. 17

  3. [11]

    Feasibility of volume-of-interest (VOI) scanning technique in cone beam breast CT—a preliminary study.Med Phys., 35(8):3482–3490, 2008

    Chen L, Shaw CC, Altunbas MC, Lai CJ, Liu X, Han T, Wang T, Yang WT, and Whitman GJ. Feasibility of volume-of-interest (VOI) scanning technique in cone beam breast CT—a preliminary study.Med Phys., 35(8):3482–3490, 2008

  4. [12]

    A general framework and review of scatter correction methods in x-ray cone-beam computerized tomography

    R¨ uhrnschopf EP and Klingenbeck K. A general framework and review of scatter correction methods in x-ray cone-beam computerized tomography. part 1: scatter compensation approaches.Med Phys., 38(7):4296–4311, 2011

  5. [13]

    Accelerating Monte Carlo simulations of photon trans- port in a voxelized geometry using a massively parallel graphics processing unit

    Badal A and Badano A. Accelerating Monte Carlo simulations of photon trans- port in a voxelized geometry using a massively parallel graphics processing unit. Med Phys., 36(11):4878–4880, 2009

  6. [14]

    Monte-Carlo scatter correction for cone-beam computed tomography with limited scan field-of-view

    Bertram M, Sattel T, Hohmann S, and Wiegert J. Monte-Carlo scatter correction for cone-beam computed tomography with limited scan field-of-view. InMedical Imaging 2008: Physics of Medical Imaging, volume 6913, page 69131Y. SPIE, 2008

  7. [15]

    U-Net: Convolutional Networks for Biomedical Image Segmentation

    Ronneberger O, Fischer P, and Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. InMedical Image Computing and Computer- Assisted Intervention – MICCAI 2015, pages 234–241, 2015

  8. [16]

    Maier J, Eulig E, V¨ oth T, Knaup M, Kuntz J, Sawall S, and Kachelrieß M. Real-time scatter estimation for medical CT using the deep scatter estimation: Method and robustness analysis with respect to different anatomies, dose levels, tube voltages, and data truncation.Med Phys...

  9. [17]

    Projection-domain scatter correction for cone beam computed tomography using a residual convolutional neural network.Med Phys., 46(7):3142–3155, 2019

    Nomura Y, Xu Q, Shirato H, Shimizu S, and Xing L. Projection-domain scatter correction for cone beam computed tomography using a residual convolutional neural network.Med Phys., 46(7):3142–3155, 2019

  10. [18]

    A deep learning approach to estimate x-ray scatter in digital breast tomosynthesis: From phantom models to clinical applications.Med Phys., 50(8):4744–4757, 2023

    Pinto MC, Mauter F, Michielsen K, Biniazan R, Kappler S, and Sechopoulos I. A deep learning approach to estimate x-ray scatter in digital breast tomosynthesis: From phantom models to clinical applications.Med Phys., 50(8):4744–4757, 2023

  11. [19]

    Task-based transferable deep-learning scatter correction in cone beam computed tomography: a simulation study.J

    Cruz-Bastida JP, Moncada F, Mart ´ ınez-D´ avalos A, and Rodr ´ ıguez-Villafuerte M. Task-based transferable deep-learning scatter correction in cone beam computed tomography: a simulation study.J. Med. Imaging, 11(2):024006–024006, 2024

  12. [20]

    Deep learning architecture for scatter estimation in cone-beam computed tomography head imaging with varying field- of-measurement settings.J Med Imaging, 11(5):053501–053501, 2024

    Agrawal H, Hietanen A, and S¨ arkk¨ a S. Deep learning architecture for scatter estimation in cone-beam computed tomography head imaging with varying field- of-measurement settings.J Med Imaging, 11(5):053501–053501, 2024

  13. [21]

    Deep learning-based forward and cross-scatter correction in dual-source CT

    Erath J, V¨ oth T, Maier J, Fourni´ e E, Petersilka M, Stierstorfer K, and Kachelrieß M. Deep learning-based forward and cross-scatter correction in dual-source CT. Med Phys., 48(9):4824–4842, 2021

  14. [22]

    Utilizing U-Net architec- tures with auxiliary information for scatter correction in CBCT across different field-of-view settings

    Harshit Agrawal, Ari Hietanen, and Simo S¨ arkk¨ a. Utilizing U-Net architec- tures with auxiliary information for scatter correction in CBCT across different field-of-view settings. In Rebecca Fahrig, John M. Sabol, and Ke Li, editors, Medical Imaging 2024: Physics of Medical...

  15. [23]

    Effect of the pixel interpolation method for downsampling medical images on deep learning accuracy.J Comput Commun., 9(11):150–156, 2021

    Hirahara D, Takaya E, Kadowaki M, Kobayashi Y, and Ueda T. Effect of the pixel interpolation method for downsampling medical images on deep learning accuracy.J Comput Commun., 9(11):150–156, 2021

  16. [24]

    The impact of downsampling methods 18 on face recognition in electronic identity card

    Hindratno MN, Nisa A, Rohim MIA, Fajri R, Hamdani M, Wibowanto GS, Les- triandoko NH, and Normakristagaluh P. The impact of downsampling methods 18 on face recognition in electronic identity card. In2023 International Confer- ence on Computer, Control, Informatics and its Appl...

  17. [25]

    A deep learning-based scatter correction of simulated x-ray images.Electronics, 8(9):944, 2019

    Lee H and Lee J. A deep learning-based scatter correction of simulated x-ray images.Electronics, 8(9):944, 2019

  18. [26]

    Evaluation of cbct scatter correction using deep convolutional neural networks for head and neck adaptive proton therapy.Phys Med Biol., 65(24):245022, 2020

    Lalonde A, Winey B, Verburg J, Paganetti H, and Sharp GC. Evaluation of cbct scatter correction using deep convolutional neural networks for head and neck adaptive proton therapy.Phys Med Biol., 65(24):245022, 2020

  19. [27]

    PENELOPE-2006: A code sys- tem for Monte Carlo simulation of electron and photon transport

    Salvat F, Fern´ andez-Varea JM, Sempau J, et al. PENELOPE-2006: A code sys- tem for Monte Carlo simulation of electron and photon transport. InWorkshop proceedings, volume 4, page 7. Citeseer, 2006

  20. [28]

    Deep Learning Based Projection Domain Metal Segmentation for Metal Artifact Reduction in Cone Beam Computed Tomography.IEEE Access, 11:100371–100382, 2023

    Agrawal H, Hietanen A, and S¨ arkk¨ a S. Deep Learning Based Projection Domain Metal Segmentation for Metal Artifact Reduction in Cone Beam Computed Tomography.IEEE Access, 11:100371–100382, 2023

  21. [29]

    Head-and-neck squa- mous cell carcinoma patients with CT taken during pre-treatment, mid- treatment, and post-treatment (HNSCC-3DCT-RT) [Dataset]

    T Bejarano, M Ornelas-Couto, and IB Mihaylov. Head-and-neck squa- mous cell carcinoma patients with CT taken during pre-treatment, mid- treatment, and post-treatment (HNSCC-3DCT-RT) [Dataset]. Online, 2018. [doi:10.7937/K9/TCIA.2018.13upr2xf]

  22. [30]

    The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository.J Digit Imaging, 26(6):1045–1057, 2013

    Clark K, Vendt B, Smith K, Freymann J, Kirby J, Koppel P, Moore S, Phillips S, Maffitt D, Pringle M, et al. The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository.J Digit Imaging, 26(6):1045–1057, 2013

  23. [31]

    Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation

    Ansel J, Yang E, He H, Gimelshein N, Jain A, Voznesensky M, Bao B, Bell P, Berard D, Burovski E, et al. Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. InProceedings of the 29th ACM International Conference on Architectu...

  24. [32]

    Nearest neighbor value interpolation.arXiv preprint arXiv:1211.1768, 2012

    Rukundo O and Cao H. Nearest neighbor value interpolation.arXiv preprint arXiv:1211.1768, 2012

  25. [33]

    Area-based interpolation for scaling of images from a CCD

    Wong PW and Herley C. Area-based interpolation for scaling of images from a CCD. InProceedings of International Conference on Image Processing, volume 1, pages 905–908, 1997

  26. [34]

    Survey: Interpolation methods in medical image processing.IEEE Trans Med Imaging, 18(11):1049–1075, 1999

    Lehmann TM, Gonner C, and Spitzer K. Survey: Interpolation methods in medical image processing.IEEE Trans Med Imaging, 18(11):1049–1075, 1999

  27. [35]

    Cubic convolution interpolation for digital image processing.IEEE Trans Acoust Speech Signal Process., 29(6):1153–1160, 1981

    Keys R. Cubic convolution interpolation for digital image processing.IEEE Trans Acoust Speech Signal Process., 29(6):1153–1160, 1981

  28. [36]

    Practical cone-beam algorithm.J Opt Soc Am A 1, 1(6):612–619, 1984

    Lee A Feldkamp, Lloyd C Davis, and James W Kress. Practical cone-beam algorithm.J Opt Soc Am A 1, 1(6):612–619, 1984. 19

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.