pith. sign in

arxiv: 2607.00270 · v1 · pith:LPBKQ2DEnew · submitted 2026-06-30 · ⚛️ physics.ins-det · cs.LG

Computer vision-based neural networks for radioisotope identification in urban environments

Masen Bachleda , Peter Lalor This is my paper

Pith reviewed 2026-07-02 16:25 UTC · model grok-4.3

classification ⚛️ physics.ins-det cs.LG
keywords radioisotope identificationconvolutional neural networkwaterfall spectrogramurban searchgamma raymachine learninganomaly detectionRADAI
0
0 comments X

The pith

Treating consecutive gamma-ray spectra as image channels enables CNNs to outperform NMF in urban radioisotope identification at low false positive rates.

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

The paper demonstrates that converting mobile gamma-ray list-mode data into waterfall spectrograms, with consecutive time spectra serving as multiple input channels, allows computer vision neural networks to identify radioisotopes amid urban backgrounds more effectively than prior methods. This tackles issues of non-uniform backgrounds, short encounters, and imbalanced data in city search operations. Evaluated on the RADAI dataset, the convolutional neural network achieves higher rates of true detection, classification, and identification than non-negative matrix factorization when the false positive rate is kept below one per hour. If correct, this indicates that adapting image-based machine learning can improve radiation detection systems for real-world mobile scenarios.

Core claim

The central claim is that the multi-channel waterfall spectrogram representation permits neural network architectures to learn distinctions between source signatures and background fluctuations better than non-negative matrix factorization, with the CNN delivering true detection, classification, and identification rates of 0.4334, 0.3965, and 0.2950 at a false positive rate below one alarm per hour, exceeding the NMF rates of 0.4151, 0.3611, and 0.2625.

What carries the argument

The waterfall spectrogram representation in which consecutive time spectra form separate input channels, analogous to color channels in images, allowing convolutional networks to process spectral and temporal patterns jointly.

If this is right

  • The CNN outperforms NMF on all global metrics at the specified false positive rate.
  • At stricter false positive constraints, neural networks perform comparably but lower than NMF.
  • The approach applies to three architectures: MLP, CNN, and ViT on the RADAI benchmark.
  • Further research is needed to improve performance at lower false positive rates.

Where Pith is reading between the lines

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

  • This channel-based representation could be tested on other time-resolved spectral datasets beyond radiation.
  • Integration with additional sensor inputs might enhance overall system robustness in urban settings.
  • The performance gains suggest potential for deployment in operational mobile search equipment if validated on diverse datasets.

Load-bearing premise

The chosen waterfall spectrogram representation with consecutive time spectra as input channels sufficiently encodes the temporal and spectral distinctions needed to separate source signatures from non-uniform urban backgrounds on the RADAI dataset.

What would settle it

Repeating the evaluation on the RADAI dataset but with single-channel spectrogram inputs instead of multi-channel would show if the channel representation is essential for the reported gains.

Figures

Figures reproduced from arXiv: 2607.00270 by Masen Bachleda, Peter Lalor.

Figure 1
Figure 1. Figure 1: Example waterfall spectrograms showing (a) background radiation and (b) a highly enriched uranium (HEU) source [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
read the original abstract

Algorithm development for radioisotope identification in mobile urban search scenarios face significant challenges from non-uniform backgrounds, momentary source encounters, and severe class imbalance between rare threat signatures and background measurements. We present a machine learning-based approach to this problem that converts list-mode gamma-ray data into two-dimensional waterfall spectrograms and applies computer vision architectures to the resulting images. Rather than treating waterfalls as conventional images, we employ a representation where consecutive time spectra can form input channels, similar to RGB channels in color images. This representation encodes both spectral and temporal information, enabling neural networks to more effectively learn patterns that distinguish source signatures from background fluctuations. We evaluate three architectures, a multilayer perceptron (MLP), convolutional neural network (CNN), and vision transformer (ViT), on the Radiological Anomaly Detection and Identification (RADAI) benchmark dataset. At a false positive rate of less than one false alarm per hour, our CNN outperforms the previous-best non-negative matrix factorization (NMF) method across all global metrics, achieving true detection, classification, and identification rates of 0.4334, 0.3965, and 0.2950 respectively, compared to 0.4151, 0.3611, and 0.2625 for NMF. At lower false positive rate constraints, the neural network approaches show comparable but ultimately lower performance than NMF, indicating opportunities for further research.

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 / 1 minor

Summary. The paper claims that converting list-mode gamma-ray data into two-dimensional waterfall spectrograms—with consecutive time spectra stacked as input channels—enables computer vision architectures (MLP, CNN, ViT) to outperform non-negative matrix factorization (NMF) for radioisotope detection, classification, and identification on the RADAI benchmark, specifically reporting CNN true rates of 0.4334/0.3965/0.2950 versus NMF rates of 0.4151/0.3611/0.2625 at false-positive rate <1 per hour.

Significance. If the results hold after additional validation, the work supplies a concrete empirical comparison against a named external baseline on a public benchmark, demonstrating that channel-stacked spectrogram inputs can yield measurable gains in a challenging mobile urban-search setting with non-uniform backgrounds and class imbalance. This provides a reproducible reference point for future instrumentation and detection studies.

major comments (2)
  1. [Abstract] Abstract: the headline performance margin is presented as arising from the CNN applied to the channel-stacked waterfall representation, yet the text supplies neither an ablation comparing this encoding to alternatives (e.g., single-spectrum inputs or explicit temporal modeling) nor any analysis showing that source-induced channel patterns are distinguishable from urban background fluctuations on the same timescale; without such evidence the reported 0.0325 absolute gain in identification rate cannot be attributed to the network rather than the data representation.
  2. [Abstract] Abstract: the three global metrics are reported to three or four decimal places with no error bars, confidence intervals, or statistical significance tests against the NMF baseline, and no training protocol, hyperparameter values, or cross-validation procedure is described; these omissions make it impossible to judge whether the claimed superiority at FPR <1/h is robust or reproducible.
minor comments (1)
  1. The abstract states that three architectures were evaluated but reports quantitative results only for the CNN; the full manuscript should clarify whether MLP and ViT results appear in a table or supplementary material and whether they also exceed NMF at the operating point of interest.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful review and constructive feedback on our manuscript. We address each major comment below, indicating the revisions planned for the next version.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the headline performance margin is presented as arising from the CNN applied to the channel-stacked waterfall representation, yet the text supplies neither an ablation comparing this encoding to alternatives (e.g., single-spectrum inputs or explicit temporal modeling) nor any analysis showing that source-induced channel patterns are distinguishable from urban background fluctuations on the same timescale; without such evidence the reported 0.0325 absolute gain in identification rate cannot be attributed to the network rather than the data representation.

    Authors: We agree that the current manuscript lacks an ablation study isolating the contribution of the channel-stacked representation and does not analyze distinguishability of source patterns from background. In the revised version we will add an ablation comparing multi-channel waterfall inputs against single-spectrum inputs and explicit temporal models, plus a feature analysis demonstrating that channel-wise source signatures differ from background fluctuations on the relevant timescales. This will strengthen attribution of the observed gain. revision: yes

  2. Referee: [Abstract] Abstract: the three global metrics are reported to three or four decimal places with no error bars, confidence intervals, or statistical significance tests against the NMF baseline, and no training protocol, hyperparameter values, or cross-validation procedure is described; these omissions make it impossible to judge whether the claimed superiority at FPR <1/h is robust or reproducible.

    Authors: The manuscript as submitted indeed omits error estimates, significance testing, and full experimental details. We will revise to report confidence intervals or standard errors on all metrics, include statistical significance tests versus the NMF baseline, and provide a complete description of the training protocol, hyperparameter values, and cross-validation procedure. These additions will allow assessment of robustness and reproducibility. revision: yes

Circularity Check

0 steps flagged

Empirical ML evaluation on external benchmark with no self-referential reductions

full rationale

The paper reports standard supervised training of MLP/CNN/ViT models on waterfall spectrograms derived from list-mode data, followed by direct comparison of detection/classification/identification rates against an independently published NMF baseline on the public RADAI dataset. No equations, parameter fits, or definitions are shown that would make the reported true-positive rates equivalent to quantities constructed from the same model; the waterfall channel-stacking is an input representation choice, not a self-defining loop. The central claim therefore rests on external data and an external comparator rather than reducing to its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; the approach rests on standard supervised learning assumptions for neural network training on image-like data.

pith-pipeline@v0.9.1-grok · 5780 in / 1066 out tokens · 36200 ms · 2026-07-02T16:25:05.955948+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

29 extracted references · 23 canonical work pages · 2 internal anchors

  1. [1]

    G. F. Knoll, Radiation Detection and Measurement, 4th Edition, John Wiley & Sons, Hoboken, NJ, 2010

    2010

  2. [2]

    K. J. Bilton, T. H. Y. Joshi, M. S. Bandstra, J. C. Curtis, D. Hellfeld, K. Vetter, Neural network approaches for mobile spectroscopic gamma-ray source detection, Journal of Nuclear Engineering 2 (2) (2021) 190–206.doi:10.3390/jne2020018

    work page doi:10.3390/jne2020018 2021

  3. [3]

    J. M. Ghawaly Jr, D. E. Peplow, D. E. Archer, T. H. Y. Joshi, M. S. Bandstra, B. J. Quiter, J. C. Curtis, A. D. Nicholson, Design and generation of a dataset for radiation detection algorithm develop- ment, Available athttps://bdc.lbl.gov/wiki/public/radai-interactive-datasets/, registration required for access (2025)

    2025

  4. [4]

    J. M. Ghawaly Jr, A. D. Nicholson, D. E. Peplow, C. M. Anderson-Cook, K. L. Myers, D. E. Archer, M. J. Willis, B. J. Quiter, Data for training and testing radiation detection algorithms in an urban environment, Scientific Data 7 (2020) 328.doi:10.1038/s41597-020-00672-2

    work page doi:10.1038/s41597-020-00672-2 2020

  5. [5]

    Kamuda, J

    M. Kamuda, J. Zhao, K. Huff, A comparison of machine learning methods for automated gamma-ray spectroscopy, Nuclear Instruments and Methods in Physics Research Section A 954 (2020) 161385. doi:10.1016/j.nima.2018.10.063

    work page doi:10.1016/j.nima.2018.10.063 2020

  6. [6]

    Daniel, F

    G. Daniel, F. Ceraudo, O. Limousin, D. Maier, A. Meuris, Automatic and real-time identification of radionuclides in gamma-ray spectra: A new method based on convolutional neural network trained with synthetic data set, IEEE Transactions on Nuclear Science 67 (4) (2020) 644–653.doi:10.1109/TNS. 2020.2969703

    work page doi:10.1109/tns 2020

  7. [7]

    Gomez-Fernandez, W

    M. Gomez-Fernandez, W. K. Wong, A. Tokuhiro, K. Welter, A. M. Alhawsawi, H. Yang, et al., Iso- tope identification using deep learning: An explanation, Nuclear Instruments and Methods in Physics Research Section A 988 (2021) 164925.doi:10.1016/j.nima.2020.164925

    work page doi:10.1016/j.nima.2020.164925 2021

  8. [8]

    A. D. Nicholson, D. E. Peplow, J. M. Ghawaly Jr, M. J. Willis, D. E. Archer, Generation of synthetic data for a radiation detection algorithm competition, IEEE Transactions on Nuclear Science 67 (8) (2020) 1968–1980.doi:10.1109/TNS.2020.3001754

    work page doi:10.1109/tns.2020.3001754 2020

  9. [9]

    Olmos, J

    P. Olmos, J. C. Diaz, J. M. Perez, P. Gomez, V. Rodellar, P. Aguayo, A. Bru, G. Garcia-Belmonte, J. L. de Pablos, A new approach to automatic radiation spectrum analysis, in: IEEE Transactions on Nuclear Science, Vol. 38, 1991, pp. 971–975.doi:10.1109/23.83860

    work page doi:10.1109/23.83860 1991

  10. [10]

    P. E. Keller, R. T. Kouzes, Gamma spectral analysis via neural networks, in: Proceedings of 1994 IEEE Nuclear Science Symposium, Vol. 1, 1994, pp. 341–345.doi:10.1109/NSSMIC.1994.474365

    work page doi:10.1109/nssmic.1994.474365 1994

  11. [11]

    Yoshida, K

    E. Yoshida, K. Shizuma, S. Endo, T. Oka, Application of neural networks for the analysis of gamma- ray spectra measured with a Ge spectrometer, Nuclear Instruments and Methods in Physics Research Section A 484 (1) (2002) 557–563.doi:10.1016/S0168-9002(01)01962-3

    work page doi:10.1016/s0168-9002(01)01962-3 2002

  12. [12]

    M. E. Medhat, Artificial intelligence methods applied for quantitative analysis of natural radioactive sources, Annals of Nuclear Energy 45 (2012) 73–79.doi:10.1016/j.anucene.2012.02.013

    work page doi:10.1016/j.anucene.2012.02.013 2012

  13. [13]

    M. J. Aitkenhead, M. Owen, D. M. Chambers, Use of artificial neural networks in measuring charac- teristics of shielded plutonium for arms control, Journal of Analytical Atomic Spectrometry 27 (2012) 432–439.doi:10.1039/C2JA10230G

    work page doi:10.1039/c2ja10230g 2012

  14. [14]

    Sheinfeld, S

    M. Sheinfeld, S. Levinson, I. Orion, Highly accurate prediction of specific activity using deep learning, Applied Radiation and Isotopes 130 (2017) 115–120.doi:10.1016/j.apradiso.2017.09.023

    work page doi:10.1016/j.apradiso.2017.09.023 2017

  15. [15]

    Kamuda, C

    M. Kamuda, C. J. Sullivan, An automated isotope identification and quantification algorithm for isotope 13 mixtures in low-resolution gamma-ray spectra, Radiation Physics and Chemistry 155 (2019) 281–286. doi:10.1016/j.radphyschem.2018.06.017

    work page doi:10.1016/j.radphyschem.2018.06.017 2019

  16. [16]

    Chen, Y.-X

    L. Chen, Y.-X. Wei, Nuclide identification algorithm based on K-L transform and neural networks, Nuclear Instruments and Methods in Physics Research Section A 598 (2) (2009) 450–453.doi:10. 1016/j.nima.2008.09.035

    2009

  17. [17]

    Lalor, H

    P. Lalor, H. Adams, A. Hagen, Sim-to-real supervised domain adaptation for radioisotope identification, Nuclear Instruments and Methods in Physics Research Section A 1083 (2026) 171159.doi:10.1016/ j.nima.2025.171159

    work page arXiv 2026

  18. [18]

    Unsupervised domain adaptation for radioisotope identification in gamma spectroscopy

    P. Lalor, A. Panigrahy, A. Hagen, Unsupervised domain adaptation for radioisotope identification in gamma spectroscopy (2026).arXiv:2603.05719. URLhttps://arxiv.org/abs/2603.05719

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  19. [19]

    J. M. Ghawaly, A. D. Nicholson, D. E. Archer, M. J. Willis, I. Garishvili, B. Longmire, et al., Char- acterization of the autoencoder radiation anomaly detection (arad) model, Engineering Applications of Artificial Intelligence 111 (2022) 104761.doi:10.1016/j.engappai.2022.104761

    work page doi:10.1016/j.engappai.2022.104761 2022

  20. [20]

    D. E. Archer, D. E. Hornback, J. O. Johnson, T. M. Miller, A. D. Nicholson, B. W. Patton, D. E. Peplow, B. Ayaz-Maierhofer, Systematic assessment of neutron and gamma backgrounds relevant to operational modeling and detection technology implementation, Tech. Rep. ORNL/TM-2014/687, Oak Ridge National Laboratory (2015).doi:10.2172/1185844

    work page doi:10.2172/1185844 2014

  21. [21]

    Lei, Robust detection of radiation threat, IEEE Nuclear Science Symposium

    E. Lei, Robust detection of radiation threat, IEEE Nuclear Science Symposium. Slides:https:// ericclei.github.io/publications/nss16-windows-slides.pdf(2016)

    2016

  22. [22]

    K. J. Bilton, T. H. Y. Joshi, M. S. Bandstra, J. C. Curtis, B. J. Quiter, R. J. Cooper, K. Vetter, Non- negative matrix factorization of gamma-ray spectra for background modeling, detection, and source identification, IEEE Transactions on Nuclear Science 66 (5) (2019) 827–837.doi:10.1109/TNS.2019. 2907267

    work page doi:10.1109/tns.2019 2019

  23. [23]

    D. D. Lee, H. S. Seung, Learning the parts of objects by non-negative matrix factorization, Nature 401 (1999) 788–791.doi:10.1038/44565

    work page doi:10.1038/44565 1999

  24. [24]

    T.-Y. Lin, P. Goyal, R. Girshick, K. He, P. Dollár, Focal loss for dense object detection, in: Proceedings of the IEEE International Conference on Computer Vision, 2017, pp. 2980–2988.doi:10.1109/ICCV. 2017.324

    work page doi:10.1109/iccv 2017

  25. [25]

    M. S. Bandstra, J. C. Curtis, J. M. Ghawaly Jr, A. C. Jones, T. H. Y. Joshi, Explaining machine- learning models for gamma-ray detection and identification, PLoS ONE 18 (6) (2023) e0286829.doi: 10.1371/journal.pone.0286829

    work page doi:10.1371/journal.pone.0286829 2023

  26. [26]

    J. L. Ba, J. R. Kiros, G. E. Hinton, Layer normalization, arXiv preprint arXiv:1607.06450 (2016). URLhttps://arxiv.org/abs/1607.06450

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  27. [27]

    Ioffe, C

    S. Ioffe, C. Szegedy, Batch normalization: Accelerating deep network training by reducing internal covariate shift, in: International Conference on Machine Learning, PMLR, 2015, pp. 448–456. URLhttps://proceedings.mlr.press/v37/ioffe15.html

    2015

  28. [28]

    Heigold, S

    A.Dosovitskiy, L.Beyer, A.Kolesnikov, D.Weissenborn, X.Zhai, T.Unterthiner, M.Dehghani, M.Min- derer, G. Heigold, S. Gelly, et al., An image is worth 16x16 words: Transformers for image recognition at scale, in: International Conference on Learning Representations, 2021. URLhttps://openreview.net/forum?id=YicbFdNTTy

    2021

  29. [29]

    Akiba, S

    T. Akiba, S. Sano, T. Yanase, T. Ohta, M. Koyama, Optuna: A next-generation hyperparameter opti- mization framework, in: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge 14 Discovery & Data Mining, 2019, pp. 2623–2631.doi:10.1145/3292500.3330701. 15 Appendix A. Hyperparameter Search Spaces This appendix provides the complete hyperp...

    work page doi:10.1145/3292500.3330701 2019