Reconstruction of Compton Edges in Plastic Gamma Spectra Using Deep Autoencoder

Byoungil Jeon, Youhan Lee, Myungkook Moon, Jongyul Kim, Gyuseong ChoSensors 2020 · OpenReview

TL;DR

This paper introduces a deep autoencoder model capable of reconstructing Compton edges in plastic gamma spectra, overcoming the inherent limitations of these detectors like poor energy resolution and the absence of photo peaks. This advancement enables direct pseudo gamma spectroscopy, even with low-count statistics, by accurately identifying gamma ray energies.

What Problem Does This Paper Address?

Plastic scintillation detectors are cost-effective and widely used in radiation monitoring due to their durability and ease of manufacturing. However, they suffer from significant energy broadening and lack distinct photo peaks, making precise radioisotope identification challenging. Existing methods for "pseudo gamma spectroscopy" often provide indirect information or require high counting statistics, limiting their practical application. This paper aims to address this by developing a method to directly reconstruct the characteristic "Compton edges" in these spectra, which can then be used to infer the incident gamma ray energy.

Methods

The core of this work is a deep autoencoder, a type of artificial neural network designed for unsupervised learning. The autoencoder consists of an encoder that compresses the input data into a lower-dimensional representation and a decoder that reconstructs the original data from this compressed representation. In this study, the autoencoder was trained to take a "broadened" (simulated with energy broadening) plastic gamma spectrum as input and reconstruct an "ideal" (simulated without broadening) spectrum. This process effectively learns to "undo" the energy broadening effect. The model was trained using a Mean Absolute Percentage Error (MAPE) as the loss function to compare the reconstructed and ideal spectra, emphasizing relative differences which are crucial for spectra with varying count levels.

Datasets

The researchers generated their own datasets for training, validation, and testing. These datasets were created using Monte Carlo simulations (MCNP 6.2) of plastic gamma spectra. To mimic real-world scenarios and address the challenge of low counting statistics, they developed a sophisticated data generation procedure. This involved:

  • Creating libraries of probability density functions (PDFs) for ideal and energy-broadened spectra of various radioisotopes (22Na, 60Co, 133Ba, 137Cs).
  • Synthesizing new spectra by combining these PDFs with varying ratios, simulating mixtures of radioisotopes.
  • Randomly sampling the number of counts for each spectrum to introduce statistical uncertainties, mimicking spectra with different counting statistics. This approach allowed them to generate a large and diverse dataset (60,000 training, 2000 validation, 2000 test spectra) that could train the autoencoder to handle low-count scenarios.

Key Insights and Ideas

  • Compton Edges as Spectroscopic Signatures: The paper highlights that even without photo peaks, the "Compton edges" in plastic scintillator spectra contain enough information to reconstruct incident gamma ray energies.
  • Deep Autoencoders for De-broadening: A deep autoencoder can effectively learn to reverse the energy broadening effect inherent in plastic scintillators, transforming a broad spectrum into a sharper, more informative one.
  • Data Generation for Robustness: The novel dataset generation procedure, which incorporates varying synthesis ratios and random sampling of counts, is crucial for training a model that performs well even with low counting statistics. This is a significant step beyond traditional methods that struggle in such conditions.
  • Minimum Reconstructible Counts (MRC): The study quantifies the minimum number of counts required for reliable Compton edge reconstruction for individual isotopes, providing practical guidance for detector usage.

Key Results

  • The trained deep autoencoder successfully reconstructed Compton edges in simulated spectra, achieving an average test loss (MAPE) of 20.019%.
  • Crucially, the model also performed well on experimentally measured spectra, demonstrating its ability to reconstruct Compton edges even when dealing with statistical uncertainties and poor counting statistics.
  • The Minimum Reconstructible Counts (MRCs) were evaluated for single isotopes, with values ranging from 650 counts for 60Co to 3750 counts for 133Ba, indicating the detector's sensitivity limits.
  • The reconstructed Compton edges closely matched theoretical values, validating the model's accuracy.

Limitations and Future Work

The authors acknowledge three main limitations:

  1. Data Specificity: The autoencoder generates results specific to the training data. If presented with spectra from radioisotopes not included in the training set, it might produce results resembling trained isotopes or mixtures.
  2. Increased MRCs with Isotope Complexity: The presence of more radioisotopes or complex spectral features (like those in 133Ba) can increase the minimum counts required for accurate reconstruction.
  3. Bare Source Assumption: The current study uses spectra from bare sources. In real-world applications, surrounding materials can distort spectra, potentially affecting Compton edge reconstruction performance.

Future work could focus on expanding the training dataset to include a wider range of radioisotopes and exploring methods to mitigate the impact of surrounding materials on spectral data.

Summary updated

Abstract

Plastic scintillation detectors are widely utilized in radiation measurement because of their unique characteristics. However, they are generally used for counting applications because of the energy broadening effect and the absence of a photo peak in their spectra. To overcome their weaknesses, many studies on pseudo spectroscopy have been reported, but most of them have not been able to directly identify the energy of incident gamma rays. In this paper, we propose a method to reconstruct Compton edges in plastic gamma spectra using an artificial neural network for direct pseudo gamma spectroscopy. Spectra simulated using MCNP 6.2 software were used to generate training and validation sets. Our model was trained to reconstruct Compton edges in plastic gamma spectra. In addition, we aimed for our model to be capable of reconstructing Compton edges even for spectra having poor counting statistics by designing a dataset generation procedure. Minimum reconstructible counts for single isotopes were evaluated with metric of mean averaged percentage error as 650 for 60Co, 2000 for 137Cs, 3050 for 22Na, and 3750 for 133Ba. The performance of our model was verified using the simulated spectra measured by a PVT detector. Although our model was trained using simulation data only, it successfully reconstructed Compton edges even in measured gamma spectra with poor counting statistics.
BibTeX
@inproceedings{jeon2020reconstruction,
  author = {Byoungil Jeon and Youhan Lee and Myungkook Moon and Jongyul Kim and Gyuseong Cho},
  title = {{Reconstruction of Compton Edges in Plastic Gamma Spectra Using Deep Autoencoder}},
  booktitle = {Sensors},
  year = {2020},
  url = {https://openreview.net/forum?id=yc917vEmo2},
}
Figure