Validation of a Bayesian-based isotope identification algorithm

Application of the Approximate Bayesian Computation Algorithm to Gamma-Ray Spectroscopy

“…There is reason to believe that RIID performance using, for example, Sodium Iodide (NaI) detectors (a very common fielded gamma ray detector) can be improved [7][8][9][10][11]. This paper will show the strong potential of approximate Bayesian computation (ABC) in peak location and area estimation, plus in simple RIID applications using either peak information or other features of the spectra collected by NaI detectors.…”

“…Since [7], several RIID advances have been made. For example, [8] introduced a Bayesian approach applied to extracted peak locations and area ratios; [9] used convolution neural networks applied to simulated spectra; [10] applied the least absolute shrinkage and selection operator (LASSO) that is well known to be effective for subset selection. The first goal of an RIID algorithm is to choose which subset from a library of a few 10 s to approximately 100 radioisotopes (including medical, threat, industrial, and naturally occurring radioactive materials) is likely to be contained in a test item.…”

“…Other radioisotopes have different gamma birth energies with different relative intensities so in theory, the potential signature of most radioisotopes is at least moderately distinct from that of other isotopes. In practice, for relatively low resolution detector types such as NaI detectors (commonly deployed detector and the main type of interest here, although higher resolution is also discussed), RIID performance can be surprisingly poor [1][2][3][4][5][7][8][9][10][15][16][17].…”

“…Since the 2009 paper [7] on RIID algorithms, to infer what radioisotopes are present in an unknown test item, there have been several improvements to RIIDs for NaI spectra. Most notably, among those algorithms that rely on estimated peak locations and net areas, [8] developed a Bayesian approach, and [10] applied the LASSO. However, reference [8] did not address to what extent its Bayesian approach actually led to well calibrated probabilities that a given radioisotope was present in each test item.…”

“…Most notably, among those algorithms that rely on estimated peak locations and net areas, [8] developed a Bayesian approach, and [10] applied the LASSO. However, reference [8] did not address to what extent its Bayesian approach actually led to well calibrated probabilities that a given radioisotope was present in each test item. That is, among all test items for which the estimated posterior probability that isotope A was present is for example, approximately 90%, was the actual fraction that included A close to 90% [25][26][27]?…”