Advanced Radiation Imaging Algorithms with Rotating Scatter Masks

The maximum-likelihood expectation-maximization (ML-EM) algorithm is frequently used for radiation image reconstruction in positron emission tomography and coded-aperture imaging devices. This algorithm may also be extended to simple source imaging through rotating scatter masks (RSMs), allowing for near full-field imaging capabilities similar to Compton cameras but at significantly higher detection efficiencies. However, degeneracies in the RSM response often cause ML-EM to fail for more complex, disjoint source distributions. Two alternate algorithms are proposed in this study in lieu of ML-EM for RSM imaging. First, the locally competitive algorithm (LCA) is adapted to model the RSM as a compressed sensor, where the signal represents a sparse code approximation of the image. Second, a convolutional neural network (CNN) is developed that upsamples the RSM signal into the original source image through an automated encoder-decoder. Monte-Carlo N-Particle Transport (MCNP) and GEometry ANd Tracking (Geant4) models of the RSM responses are integrated into both algorithms. This integration marks a significant deviation from standard neural networks and machine learning techniques, often implemented as black boxes which ignore any a priori knowledge of the processes involved. Performance of the three algorithms are systematically evaluated over a range of simulated source distributions as well as experimentally-collected RSM signals for multiple point sources.