Deep Learning–Based Image Reconstruction of Coded-Aperture Imaging in Nuclear Security Applications

Fast and accurate localization of radioactive sources is critical to the health and safety of workers in nuclear security applications. Coded-aperture imaging is an attractive collimator-based technique for the remote localization of radioactive sources because of its excellent angular resolution and sensitivity but requires an important image reconstruction step. Algorithms with varying computational complexity and performance have been proposed. An analytical algorithm such as cross-correlation provides high computational efficiency but with low resolution and great background noise. It can also cause artifacts when the sources are located at the partially coded field of view (PCFOV) for the MURA mask-based imager. An iterative algorithm such as MLEM offers great reconstruction quality but with high computational cost. In this chapter, we discuss the detection and localization method of radioactive sources of a coded-aperture imager in nuclear security applications. The theory of coded-aperture imaging and image reconstruction algorithms are introduced. A fast and accurate image reconstruction method based on deep learning is proposed and evaluated using both simulation and experimental data. The model is trained on Monte Carlo simulation data with different conditions, and data augmentation is utilized to improve the data acquisition efficiency. Both the fully coded field if view (FCFOV) and the PCFOV are considered to extend the imaging area of the coded-aperture imager.