Fig 1.
Example of masking a gamma-ray spectrum (black) using different approaches mentioned in the text. The spectrum contains an 131I anomaly, whose main photopeak can be seen at 364 keV.
Fig 2.
The effect of the masking methods on model class scores.
Examples of the distribution of the model’s class scores for 131I using different masking methods. At top are the class scores for 131I for a background spectrum, and the bottom shows the same quantities, but for the spectrum shown in Fig 1, which contains 131I. Each histogram contains the results of 20,000 random masks.
Table 1.
Anomaly types in the RADAI dataset.
Table 2.
Model hyperparameter search.
Table 3.
Model architecture.
Fig 3.
Confusion matrix of the model.
The confusion matrix for the 25 spectrum categories obtained by evaluating the model on the testing set.
Fig 4.
Comparison of explanation methods for a spectrum with 137Cs.
A comparison of the four explanation methods considered here for a spectrum containing a medium energy source, 137Cs. For each spectrum, the scaled mean background spectrum is shown for reference, and the deviance residual from the scaled mean background is shown beneath.
Fig 5.
Comparison of explanation methods for a spectrum with 201Tl.
A comparison of the four explanation methods considered here for a spectrum with a low energy source, 201Tl. For clarity, the spectrum is shown only up to 1000 keV.
Fig 6.
Explanations for shielded and unshielded versions of the same source.
A comparison of saliency (top) and Kernel SHAP (bottom) for unshielded (left) and shielded (right) weapons-grade plutonium sources (WGPu). For clarity, the spectrum is shown only up to 1000 keV.
Fig 7.
Example of a counterfactual explanation.
Kernel SHAP used for explanations of the class score for 131I (top left), which is the true anomaly contained by the spectrum, and an explanation of the class score for 133Ba, to which the model also gave significant confidence. The bottom middle is a plot of Kernel SHAP used to generate a counterfactual explanation for why 131I was correctly selected over 133Ba.
Fig 8.
Example of a counterfactual explanation when incorrect class was identified.
Kernel SHAP used for explanations of the class score for 133Xe (top left), which is the true source present, but the model incorrectly assigned a higher confidence to 201Tl (top right). The bottom middle is a plot of the counterfactual explanation for why the incorrect 201Tl was selected instead of the correct 133Xe.