pith. sign in

arxiv: 2512.17654 · v3 · submitted 2025-12-19 · 💻 cs.LG · physics.comp-ph· physics.med-ph

Learning-Based Estimation of Spatially Resolved Scatter Radiation Fields in Interventional Radiology

Felix Lehner , Pasquale Lombardo , Susana Castillo , Oliver Hupe , Marcus Magnor This is my paper

Pith reviewed 2026-05-16 20:38 UTC · model grok-4.3

classification 💻 cs.LG physics.comp-phphysics.med-ph
keywords neural networksscatter radiationinterventional radiologydosimetryMonte Carlo simulationradiation fieldsmachine learning
0
0 comments X

The pith

Lightweight neural networks estimate three-dimensional scatter radiation fields in interventional radiology with over 84 percent accuracy on synthetic data.

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

The paper introduces three variants of lightweight fully connected neural networks for estimating three-dimensional volumes of scattered radiation fields during procedures like interventional radiology and cardiology. These networks are trained on synthetic datasets of increasing complexity generated by Monte Carlo simulations with the Alderson RANDO phantom as the primary scatter object. Evaluation across multiple metrics shows good spatial agreement between predicted and ground-truth fields, with the symmetric mean absolute percentage error for scatter radiation fields remaining consistently above 84 percent. The approach targets faster computation than full Monte Carlo runs to support dosimetry applications where scatter exposure outside the primary beam matters.

Core claim

Three variants of lightweight fully connected artificial neural networks can estimate three-dimensional, spatially resolved volumes of scattered radiation fields and their corresponding fluence and spectra distributions when trained on synthetic Monte Carlo datasets generated with RadField3D using the torso of a male Alderson RANDO phantom.

What carries the argument

Lightweight fully connected neural networks that map input parameters describing the radiation setup to full 3D fluence and spectra distributions of the scatter radiation field.

If this is right

  • The networks enable interactive, real-time estimation of radiation fields for dosimetry in interventional radiology and cardiology.
  • High accuracy in regions of interest supports practical use for out-of-field dosimetry calculations.
  • Open-source publication of the datasets and training pipeline allows others to replicate or extend the models on new configurations.

Where Pith is reading between the lines

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

  • If the networks generalize beyond the phantom data, they could be integrated into clinical software to provide on-the-fly dose estimates during procedures.
  • Such estimates might eventually guide adjustments in equipment positioning or shielding to reduce unnecessary scatter exposure to staff and patients.

Load-bearing premise

The synthetic datasets generated with the Alderson RANDO phantom and RadField3D Monte Carlo simulations accurately capture the scatter behavior that would occur with real patients and equipment in clinical interventional radiology.

What would settle it

Direct comparison of neural network predictions against physical dosimeter measurements of scatter radiation fields collected during an actual interventional radiology procedure with the same equipment and patient-like geometry.

Figures

Figures reproduced from arXiv: 2512.17654 by Felix Lehner, Marcus Magnor, Oliver Hupe, Pasquale Lombardo, Susana Castillo.

Figure 1
Figure 1. Figure 1: Distribution of relative voxel fluences inside datasets DS-01 and DS-02 for range Φ3 γ ≥ 0.0 on the X-axis, exhibiting a strong imbalance towards Φ3 γ ≤ 10−2 . 0.2 0.4 0.6 0.8 1 0 0.0005 0.001 0.0015 0.002 0.0025 0.003 0.0035 0.004 Normalized Fluence Normalized V o xel Count [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Distribution of relative voxel fluences inside dataset DS-03 for range Φ3 γ ≥ 0.0 on the X-axis, exhibiting a strong imbalance towards Φ3 γ ≤ 10−5 . 0.2 0.4 0.6 0.8 1 0 50μ 100μ 150μ 200μ Normalized Fluence Normalized V o xel Count [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Schematic of the fully connected architecture used as the backbone for SRBFNet, SPERFNet and PBRFNet outputting a spectrum and a fluence for each voxel location. Static Rotatable Beam Field Network (SRBFNet): This neural network forms the basis for reconstructing the presented datasets. For each dataset, we present variations of this network. In general, SRBFNet receives the same input parameters as the or… view at source ↗
read the original abstract

We present three variants of a lightweight, fully connected artificial neural network, suited for interactive estimation of three-dimensional, spatially resolved volumes of scattered radiation fields and a corresponding training pipeline for radiation protection dosimetry in medical radiation fields, such as those found in interventional radiology and cardiology. Accompanying, we present three different synthetically generated datasets with increasing complexity for training, generated using RadField3D, a Monte Carlo simulation application based on Geant4. As the primary scatter object, we employed the torso of a male Alderson RANDO phantom. On those datasets, we evaluate convolutional and fully connected architectures of neural networks to demonstrate which design decisions work well for reconstructing the fluence and spectra distributions over the spatial domain of such radiation fields. All our datasets, as well as our training pipeline, are published as open source in separate repositories. To evaluate the presented neural networks, we define and assess several metrics. Across these measures, the model variants demonstrate good spatial agreement between predicted and ground-truth radiation fields, particularly within specific regions of interest within the radiation field. Of particular relevance for potential application in out-of-field dosimetry is the SMAPE of the scatter radiation field, which represents the most challenging metric and was consistently above 84 %.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The manuscript presents three variants of lightweight fully connected neural networks for interactive estimation of three-dimensional spatially resolved scatter radiation fields in interventional radiology. Networks are trained and evaluated on three synthetic datasets of increasing complexity generated with the Alderson RANDO torso phantom in RadField3D (Geant4 Monte Carlo), with open-source release of data and training pipeline. Reported results show good spatial agreement between predictions and ground truth, with SMAPE consistently above 84% for scatter fields, particularly in regions of interest.

Significance. If the generalization to real clinical settings holds, the work could enable fast, interactive estimation of scatter fields as an alternative to full Monte Carlo simulations for out-of-field dosimetry. The open release of datasets and pipeline is a clear strength supporting reproducibility. The systematic comparison of network architectures on datasets of graded complexity provides useful design insights for this application domain.

major comments (2)
  1. [Abstract and Evaluation] Abstract and Evaluation sections: The claim of utility for out-of-field dosimetry in interventional radiology requires generalization beyond the training distribution, yet all experiments use only the fixed male Alderson RANDO phantom geometry with fixed table attenuation and simulated C-arm positions. No real-patient measurements, multi-phantom cross-validation, or domain-adaptation tests are reported, so the SMAPE >84% result remains specific to this single synthetic distribution.
  2. [Methods and Results] Methods and Results: The networks are fitted exclusively to Geant4 simulation outputs for one phantom; this makes the reported spatial agreement and SMAPE metrics load-bearing only within that narrow distribution. Without additional validation on variable body habitus, tissue compositions, or equipment configurations, the central applicability claim cannot be assessed from the presented evidence.
minor comments (2)
  1. [Abstract] Abstract: The text refers to evaluating both convolutional and fully connected architectures, yet the presented variants and title emphasize fully connected networks; clarify which architectures were ultimately retained and why.
  2. [Figures/Tables] Figures and tables: Regions of interest where agreement is strongest are mentioned but not explicitly delineated in the reported metrics or visualizations; adding clear ROI masks or sub-tables would improve interpretability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the need for broader validation. Our work is a controlled proof-of-concept study using synthetic data from a standardized phantom to demonstrate feasibility of lightweight networks for scatter field estimation. We address the two major comments below with partial revisions to the abstract, discussion, and limitations sections to better contextualize the scope and outline future extensions.

read point-by-point responses

  1. Referee: [Abstract and Evaluation] Abstract and Evaluation sections: The claim of utility for out-of-field dosimetry in interventional radiology requires generalization beyond the training distribution, yet all experiments use only the fixed male Alderson RANDO phantom geometry with fixed table attenuation and simulated C-arm positions. No real-patient measurements, multi-phantom cross-validation, or domain-adaptation tests are reported, so the SMAPE >84% result remains specific to this single synthetic distribution.

    Authors: We agree that the reported SMAPE metrics are specific to the synthetic distribution generated with the Alderson RANDO phantom. This phantom is the established standard for such dosimetry studies, and our three datasets systematically vary C-arm positions, field sizes, and scatter complexity while keeping the geometry fixed to enable reproducible architecture comparisons. The open-source release of data and pipeline is intended to support exactly the extensions the referee suggests. In revision we will update the abstract and add an explicit limitations paragraph stating that clinical generalization remains to be demonstrated and that the current results establish baseline performance on this reference geometry. revision: partial

  2. Referee: [Methods and Results] Methods and Results: The networks are fitted exclusively to Geant4 simulation outputs for one phantom; this makes the reported spatial agreement and SMAPE metrics load-bearing only within that narrow distribution. Without additional validation on variable body habitus, tissue compositions, or equipment configurations, the central applicability claim cannot be assessed from the presented evidence.

    Authors: The single-phantom design was chosen deliberately to isolate the effect of network architecture and training data complexity, as noted in the referee's significance assessment. Tissue compositions follow the RANDO phantom's standard specifications, and table attenuation plus C-arm geometry are modeled. We will expand the discussion to include a dedicated paragraph on applicability limits and potential transfer-learning approaches for variable habitus. No new experiments are added in this revision, but the released code base allows immediate community-driven multi-phantom tests. revision: partial

Circularity Check

0 steps flagged

No circularity: supervised learning on independent Monte Carlo data

full rationale

The paper generates three synthetic datasets via RadField3D (Geant4 Monte Carlo) using a fixed Alderson RANDO phantom, then trains fully-connected and convolutional networks to regress fluence/spectra from spatial inputs. Reported metrics (spatial agreement, SMAPE >84 %) are computed on held-out splits of the same simulation data. No equation or claim reduces to a self-definition, fitted parameter renamed as prediction, or self-citation chain; the mapping is learned from external physics simulation rather than constructed from the network itself. Generalization to real patients is an untested assumption but does not create circularity in the reported derivation.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that neural networks can learn a useful mapping from simulation inputs to full 3D fluence and spectra, plus the representativeness of the chosen phantom and Monte Carlo code.

free parameters (1)
  • neural network weights and biases
    All network parameters are fitted during training on the synthetic datasets; no specific numerical values are given in the abstract.
axioms (1)
  • domain assumption A fully connected neural network can approximate the continuous mapping from input parameters to spatially resolved radiation fluence and spectra
    Invoked by training the networks as regressors on the Monte Carlo outputs.

pith-pipeline@v0.9.0 · 5532 in / 1285 out tokens · 39668 ms · 2026-05-16T20:38:34.375482+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages

  1. [1]

    O’Connor, C

    U. O’Connor, C. Walsh, D. Gorman, G. O’Reilly, Z. Martin, P. Madhavan, R. T. Murphy, R. Szirt, A. Alm´ en, M. Andersson, A. Camp, V. Garcia, M. A. Duch, M. Ginjaume, M. Abdelrahman, P. Lombardo, and F. Vanhavere. Feasibility study of computational occupational dosimetry: evaluating a proof-of-concept in an endovascular and interventional cardiology settin...

    work page 2022

  2. [2]

    Chatzisavvas, G

    N. Chatzisavvas, G. Priniotakis, M. Papoutsidakis, D. Nikolopoulos, I. Valais, and G. Karpetas. Monte Carlo computational software and methods in radiation dosimetry. Annals of Emerging Technologies in Computing, 5(3):36–50, 06 2021

    work page 2021

  3. [3]

    Rainford, A

    L. Rainford, A. Tcacenco, J. Potocnik, C. Brophy, A. Lunney, D. Kearney, and M. O’Connor. Student perceptions of the use of three-dimensional (3-D) virtual reality (VR) simulation in the delivery of radiation protection training for radiography and medical students.Radiography, 29(4):777–785, 2023

    work page 2023

  4. [4]

    Fujiwara, S

    A. Fujiwara, S. Fujimoto, R. Ishikawa, and A. Tanaka. Virtual reality training for radiation safety in cardiac catheterization laboratories - an integrated study.Radiation Protection Dosimetry, 200(15):1462–1469, 09 2024

    work page 2024

  5. [5]

    Lehner, P

    F. Lehner, P. Lombardo, S. Castillo, O. Hupe, and M. Magnor. RadField3D: a data generator and data format for deep learning in radiation-protection dosimetry for medical applications.Journal of Radiological Protection, 45(2):021508, May 2025. Publisher: IOP Publishing

    work page 2025

  6. [6]

    Agostinelli, J

    S. Agostinelli, J. Allison, K. Amako, J. Apostolakis, and et al. Geant4—a simulation toolkit. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 506(3):250–303, 2003

    work page 2003

  7. [7]

    C. J. Werner, J. S. Bull, C. J. Solomon, F. B. Brown, G. W. McKinney, M. E. Rising, D. A. Dixon, R. L. Martz, H. G. Hughes, L. J. Cox, A. J. Zukaitis, J. C. Armstrong, R. A. Forster, and L. Casswell. MCNP Version 6.2.Release Notes, 2 2018

    work page 2018

  8. [8]

    Metrology Research Centre

    National Research Council of Canada. Metrology Research Centre. Ionizing Radiation Standards. EGSnrc: software for Monte Carlo simulation of ionizing radiation,

    work page

  9. [9]

    Record identifier: 7fbd0900-681b-4e95-95b8-b0b6082e7776 ER

    Collection: NRC Publications Archive. Record identifier: 7fbd0900-681b-4e95-95b8-b0b6082e7776 ER

    work page

  10. [10]

    S. Jan, G. Santin, D. Strul, S. Staelens, and et al. GATE: a simulation toolkit for PET and SPECT.Physics in Medicine & Biology, 49(19):4543, sep 2004

    work page 2004

  11. [11]

    Santiago Rond´ on, P

    D. Santiago Rond´ on, P. Lombardo, M. Abdelrahman, L. Struelens, F. Vanhavere, W. Schoonjans, and N. Bergans. Computational system for extremity dosimetry in nuclear medicine.Journal of Radiological Protection, 45(4):041502, oct 2025

    work page 2025

  12. [12]

    Wagner, K

    M. Wagner, K. Dresing, and W. Ludwig. SIScaR-GPU: fast simulation and visualization of intraoperative scattered radiation to support radiation protection training.Studies in Health Technology and Informatics, 180:968–972, 2012

    work page 2012

  13. [13]

    J. Wang, R. Shi, W. Zheng, W. Xie, D. Kao, and H.-N. Liang. Effect of Frame Rate on User Experience, Performance, and Simulator Sickness in Virtual Reality.IEEE Transactions on Visualization and Computer Graphics, 29(5):2478–2488, May 2023. 16 IOP PublishingJournalvv(yyyy) aaaaaa Lehneret al

    work page 2023

  14. [14]

    Hupe and U

    O. Hupe and U. Ankerhold. Determination of the dose to persons assisting when X-radiation is used in medicine, dentistry and veterinary medicine.Radiation Protection Dosimetry, 144(1-4):478–481, 11 2010

    work page 2010

  15. [15]

    Abdelrahman, P

    M. Abdelrahman, P. Lombardo, F. Vanhavere, A. Seret, C. Phillips, and P. Covens. First steps towards online personal dosimetry using computational methods in interventional radiology: Operator’s position tracking and simulation input generation.Radiation Physics and Chemistry, 171:108702, 2020

    work page 2020

  16. [16]

    Vanhavere, O

    F. Vanhavere, O. Van Hoey, P. Lombardo, M. Abdelrahman, M. Amor Duch, A. Alm´ en, U. O’Connor, R. Tanner, and E. Carinou.D9.121 – Final Report of the PODIUM project. EC - European Commission, Belgium, January 2020. Score=1

    work page 2020

  17. [17]

    Vanhavere and O

    F. Vanhavere and O. Van Hoey. Advances in personal dosimetry towards real-time dosimetry.Radiation Measurements, 158:106862, 2022

    work page 2022

  18. [18]

    Raayai Ardakani, L

    M. Raayai Ardakani, L. Yu, D. R. Kaeli, and Q. Fang. Framework for denoising Monte Carlo photon transport simulations using deep learning.Journal of Biomedical Optics, 27(8):083019, August 2022

    work page 2022

  19. [19]

    T. Bai, B. Wang, D. Nguyen, and S. Jiang. Deep dose plugin: towards real-time Monte Carlo dose calculation through a deep learning-based denoising algorithm.Machine Learning: Science and Technology, 2(2):025033, apr 2021

    work page 2021

  20. [20]

    Z. Peng, H. Shan, T. Liu, X. Pei, G. Wang, and X. G. Xu. MCDNet – a denoising convolutional neural network to accelerate Monte Carlo radiation transport simulations: A proof of principle with patient dose from X-ray CT imaging.IEEE Access, 7:76680–76689, 2019

    work page 2019

  21. [21]

    Villa Arias.Radiation exposure in x-ray guided interventions by using deep learning and Monte Carlo simulations

    M. Villa Arias.Radiation exposure in x-ray guided interventions by using deep learning and Monte Carlo simulations. Theses, Universit´ e de Bretagne occidentale - Brest, May 2023

    work page 2023

  22. [22]

    R. Ye, D. Niu, L. Li, and X. Hou. Research on a three-dimensional radiation field reconstruction algorithm based on an improved 3D CNN.Applied Radiation and Isotopes: Including Data, Instrumentation and Methods for Use in Agriculture, Industry and Medicine, 214:111540, December 2024

    work page 2024

  23. [23]

    Yisheng, W

    H. Yisheng, W. Zhen, P. Yanheng, Z. Yuhang, Q. Rui, Z. Hui, and L. Junli. Validation of the neural network for 3D photon radiation field reconstruction under various source distributions.Frontiers in Energy Research, 11, April 2023. Publisher: Frontiers

    work page 2023

  24. [24]

    Xu and L

    W. Xu and L. Bin. Research on 3D radiation fields reconstruction based on FNN.Advances in Engineering Technology Research, 2025

    work page 2025

  25. [25]

    Mildenhall, P

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis.Commun. ACM, 65(1):99–106, December 2021

    work page 2021

  26. [26]

    C ¸ i¸ cek, A

    ¨O. C ¸ i¸ cek, A. Abdulkadir, and S. S. Lienkamp. 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. In Sebastien Ourselin, Leo Joskowicz, Mert R. Sabuncu, Gozde Unal, and William Wells, editors,Medical Image Computing and Computer-Assisted Intervention – MICCAI 2016, pages 424–432, Cham, 2016. Springer International Publishing. 17 ...

    work page 2016

  27. [27]

    International Organization for Standardization. ISO 4037-1: Radiological protection — X and gamma reference radiation for calibrating dosemeters and doserate meters and for determining their response as a function of photon energy. https://www.iso.org/standard/66872.html, 2019. Last Accessed: 2024-10-23

    work page 2019

  28. [28]

    Poludniowski, A

    G. Poludniowski, A. Omar, R. Bujila, and P. Andreo. Technical Note: SpekPy v2.0 — a software toolkit for modeling X-ray tube spectra.Medical Physics, 48(7):3630–3637, 2021

    work page 2021

  29. [29]

    Hornik, M

    K. Hornik, M. Stinchcombe, and H. White. Multilayer feedforward networks are universal approximators.Neural Networks, 2(5):359–366, January 1989

    work page 1989

  30. [30]

    Loshchilov and F

    I. Loshchilov and F. Hutter. SGDR: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations, 2017

    work page 2017

  31. [31]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regularization. InInternational Conference on Learning Representations, 2019

    work page 2019

  32. [32]

    Vaswani, N

    A. Vaswani, N. Shazeer, and N. Parmar. Attention is All you Need. InAdvances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017

    work page 2017

  33. [33]

    M¨ uller, A

    T. M¨ uller, A. Evans, C. Schied, and A. Keller. Instant neural graphics primitives with a multiresolution hash encoding.ACM Trans. Graph., 41(4), July 2022

    work page 2022

  34. [34]

    M¨ uller

    T. M¨ uller. Tiny CUDA Neural Networks, 2021

    work page 2021

  35. [35]

    Perez, F

    E. Perez, F. Strub, and H. de Vries. FiLM: Visual Reasoning with a General Conditioning Layer.AAAI Conference on Artificial Intelligence, 32(1), April 2018

    work page 2018

  36. [36]

    Arevalo, T

    J. Arevalo, T. Solorio, M. Montes-y G´ omez, and F. A. Gonz´ alez. Gated multimodal networks.Neural Comput. Appl., 32(14):10209–10228, July 2020

    work page 2020

  37. [37]

    Ronneberger, P

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

    work page 2015

  38. [38]

    PyTorch Lightning, March 2025

    William Falcon and The PyTorch Lightning team. PyTorch Lightning, March 2025

    work page 2025

  39. [39]

    L. N. Smith. Cyclical learning rates for training neural networks. In2017 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 464–472, 2017

    work page 2017

  40. [40]

    Akiba, S

    T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama. Optuna: A Next-generation Hyperparameter Optimization Framework. InACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ’19, pages 2623–2631, New York, NY, USA, July 2019. Association for Computing Machinery

    work page 2019

  41. [41]

    Z. Wang, A. C. Bovik, and H. R. Sheikh. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Processing, 13(4):600–612, April 2004

    work page 2004

  42. [42]

    L. Yu, T. L. S. Tang, and N. Cassim. Analysis of dose comparison techniques for patient-specific quality assurance in radiation therapy.Journal of Applied Clinical Medical Physics, 20(11):189–198, October 2019

    work page 2019

  43. [43]

    S. B. Jiang, G. C. Sharp, and T. Neicu. On dose distribution comparison.Physics in Medicine and Biology, 51(4):759–776, February 2006. 18

    work page 2006