Deep Learning for Passive Gamma Emission Tomography

In this paper, we address the problem of generating and enhancing Passive Gamma Emission Tomography (PGET) data from a deep learning perspective. The PGET instrument has been developed for the verification of spent nuclear fuel and relies on image reconstruction and analysis algorithms to detect missing or substituted fuel pins. High quality simulations are required for validating and further improving such techniques. However, reproducing the behavior of the original instrument is not straight-forward: complex Monte Carlo simulations need to be evaluated in order to compute vast amounts of photon histories. This makes this task extremely time consuming, taking up to several days to compute a single measurement. We propose the use of Convolutional Neural Networks (CNNs) for complementing and speeding up such process. More specifically, in this work we introduce a U-Net autoencoder that learns the mapping between incomplete/noisy data and its corresponding full sinogram. To do so, we exploit the fact that sinograms are highly redundant: the contribution of a single pin can be observed from many different directions. Our CNN learns the underlying model of the data to effectively exploit these redundancies and to make informed predictions of complete PGET sinograms starting from partial views. The experimental evaluation of our trained system shows very accurate results using only a small fraction of data and with execution times below one second. The technique is suitable for efficiently generating synthetic PGET sinograms, as well as for enhancing real measurement data.