Radiation Physics and Engineering
A peer-reviewed journal published by K. N. Toosi University of Technology

Comparison of nonlinear autoencoder and linear PCA dimensionality reduction in gamma-ray spectroscopy based radioisotope identification

Document Type : Research Article

Authors

Kazem Abdolpour
S. Farhad Masoudi
Atefeh Fathi

Department of Physics, K.N. Toosi University of Technology, P.O. Box 15875-4416, Tehran, Iran

https://doi.org/10.22034/rpe.2026.580714.1354
Abstract
Dimensionality reduction can play an important role in radioisotope identification from gamma-ray spectra by compressing spectral information, reducing model complexity, and improving learning efficiency. The importance of this approach becomes more evident under real-world conditions, where spectra are typically characterized by high dimensionality, low counts, peak overlap, and calibration instabilities, all of which make direct analysis more difficult. On this basis, in the present study, two dimensionality-reduction approaches -principal component analysis as a linear method and an autoencoder as a nonlinear method -were compared to evaluate their effectiveness for radioisotope identification. Using a simulated dataset of 1024-channel NaI(Tl) spectra of six common radioisotopes (Co-60, Cs-137, I-131, Ba-133, Am-241, and Tc-99m) for training, we evaluated both approaches with a common multilayer perceptron (MLP) classifier on unseen data, including laboratory-measured spectra and scenarios with gain drift. Under ideal conditions, both approaches achieved nearly perfect identification performance (F1 ≈ 0.98-0.99). In more challenging regimes, including experimental spectra and severe gain drift, the autoencoder’s latent features consistently outperform PCA. The autoencoder+MLP model generalized better to real spectra (e.g., achieving F1 ≈ 0.99 versus 0.91 for PCA) and maintained higher accuracy under a ±20% gain drift (F1 ≈ 0.81 versus 0.71). These results suggest that the nonlinear latent representation learned by the autoencoder is less sensitive to gain drift and experimental variability than variance-based linear projections, resulting in improved robustness for multi-label radioisotope identification. This insight can support the design of more reliable field-deployable gamma spectroscopy systems, where maintaining high identification performance amid noise and calibration variability is essential.

Highlights

  • Compared PCA and autoencoder for reducing gamma-ray spectral dimensionality.
  • Trained both methods on simulated NaI(Tl) spectra and tested on real measurements.
  • Both methods achieved near-perfect isotope ID under ideal conditions.
  • Autoencoder showed higher robustness to ±20% gain drift than PCA.
  • Autoencoder features transferred better to real spectra than PCA features.

Keywords

Gamma-ray spectroscopy
Radioisotope identification
Dimensionality reduction
Autoencoder neural networks
Principal component analysis

Copyright
RPE is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

Conflict of Interest
The authors declare no potential conflict of interest regarding the publication of this work‎.

Funding
‎The authors declare that no funds‎, ‎grants‎, ‎or other financial support were received during the preparation of this manuscript‎.

Abdolpour, K., Masoudi, S. F., and Fathi, A. (2026). Using autoencoder feature compression combined with deep neural networks for precise radioisotope identification in composite spectra. Results in Physics, page 108608.
Burr, T. and Hamada, M. (2009). Radio-isotope identification algorithms for NaI γ spectra. Algorithms, 2(1):339–360.
Galib, S., Bhowmik, P., Avachat, A., et al. (2021). A comparative study of machine learning methods for automated identification of radioisotopes using NaI gamma-ray spectra. Nuclear Engineering and Technology, 53(12):4072–4079.
Goodfellow, I., Bengio, Y., Courville, A., et al. (2016). Deep learning, volume 1. MIT press Cambridge.
Hinton, G. E. and Salakhutdinov, R. R. (2006). Reducing the dimensionality of data with neural networks. Science, 313(5786):504–507.
Jolliffe, I. (2025). Principal component analysis. In International Encyclopedia of Statistical Science, pages 1945–1948. Springer.
Jolliffe, I. T. and Cadima, J. (2016). Principal component analysis: a review and recent developments. Philosophical transactions of the royal society A: Mathematical, Physical and Engineering Sciences, 374(2065):20150202.
Khatiwada, A., Klasky, M., Lombardi, M., et al. (2023). Machine Learning technique for isotopic determination of radioisotopes using HPGe γ-ray spectra. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 1054:168409.
Li, C., Liu, S., Wang, C., et al. (2022). A new radionuclide identification method for low-count energy spectra with multiple radionuclides. Applied Radiation and Isotopes, 185:110219.
Novak, R., Bahri, Y., Abolafia, D. A., et al. (2018). Sensitivity and generalization in neural networks: an empirical study. arXiv preprint arXiv:1802.08760.
Turner, A. N., Wheldon, C., Wheldon, T. K., et al. (2021). Convolutional neural networks for challenges in automated nuclide identification. Sensors, 21(15):5238.
Tutt, J., Anderson, C., and McKinney, G. (2017). Background-Source Cosmic-Photon Elevation Scaling and Cosmic-Neutron/Photon Date Scaling in MCNP6. Physics Procedia, 90:237–241.
Wang, Y., Zhang, Q., Yao, Q., et al. (2022). Multiple radionuclide identification using deep learning with channel attention module and visual explanation. Frontiers in Physics, 10:1036557.
Zhang, S., Marsden, E., and Goulermas, J. Y. (2022). Isotope identification using artificial neural network ensembles and bin-ratios. IEEE Transactions on Nuclear Science, 69(6):1194–1202.
Radiation Physics and Engineering
Volume 7, Issue 3
Summer 2026
Pages 63-73

  • Receive Date 11 May 2026
  • Revise Date 25 June 2026
  • Accept Date 28 June 2026

Home

Submit Manuscript

Reviewers

Contact Us