Restricted Boltzmann Spectrum Deconvolution

A machine learning method is shown that deconvolves spectra in a single step procedure. This method is based on a special neural network, Bernoulli type restricted Boltzmann machine (RBM). Using sophisticated Geant4 simulations to produce mono energetic line response vectors of a Sodium Iodide detec...

Full description

Saved in:
Bibliographic Details
Published in2019 IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC) pp. 1 - 3
Main Author Neuer, Marcus J.
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.10.2019
Subjects
Deconvolution
Detectors
Geant4
inversion
Machine learning
Machine learning algorithms
Neural networks
Neurons
nuclide identification
probability distribution
response matrix
restricted Boltzmann machine
spectrum enhancement
Training
Online AccessGet full text

Cover

More Information
Summary:A machine learning method is shown that deconvolves spectra in a single step procedure. This method is based on a special neural network, Bernoulli type restricted Boltzmann machine (RBM). Using sophisticated Geant4 simulations to produce mono energetic line response vectors of a Sodium Iodide detector, this RBM is trained to learn the detector response and apply the corresponding inversion to each spectrum that is inserted in the visible layer. Although training is time intensive and requires several passes through the response matrix, the evaluation of the RBM is fast. The method provides comparable results to the Maximum-Likelihood deconvolution (MLEM), but without needing iterations. Tests were performed on a series of real-world measurement spectra (not simulations), including NORM and SNM sources. The RBM was able to reconstruct the incident photon energies well. Residual noise is considerably low and the peak positions, as well, as peak height ratios are preserved well.
ISSN:2577-0829
DOI:10.1109/NSS/MIC42101.2019.9059685