Regenerative Neural Network For Rotating Scatter Mask Radiation Imaging

This paper describes the design and testing of a new form of convolutional neural network, a regenerative neural network (ReGeNN), for application to rotating scatter mask gamma imaging. The network was trained using detector responses for realistic source distributions simulated in MCNP v6.1.4. ReGeNN was shown to reconstruct the source images with excellent quality when trained under ideal conditions. When comparing to standard maximum-likelihood expectation-maximization algorithms, ReGeNN reduced the relative error from 145% to 33% and increased the precision from 27% to 85% averaged over the 24 distributed sources tested. The network also demonstrated robust learning capabilities after successfully training on noisy input data, with only relatively minor degradation to the source reconstruction quality. An analysis of variance study determined that the most significant factor affecting the reconstruction quality was the source's shape, with ring-type source distributions having the worst performance. The interaction between the source's size and direction was also discovered to have a small effect as larger sources located near the bottom of the system's field-of-view contained more phantoms within the reconstruction. Reconstruction quality was lower for responses exceeding the training noise level and for source distributions not included in the training set, indicating the importance of robust training data. The results show a significant improvement over more conventional algorithms, suggesting that real-time gamma imaging with the rotating scatter mask may be not only plausible, but practical for the first time. ReGeNN may readily be adapted for similar time-encoded radiation imaging systems, but the neural network methods described also have significant application potential towards other imaging systems.This paper describes the design and testing of a new form of convolutional neural network, a regenerative neural network (ReGeNN), for application to rotating scatter mask gamma imaging. The network was trained using detector responses for realistic source distributions simulated in MCNP v6.1.4. ReGeNN was shown to reconstruct the source images with excellent quality when trained under ideal conditions. When comparing to standard maximum-likelihood expectation-maximization algorithms, ReGeNN reduced the relative error from 145% to 33% and increased the precision from 27% to 85% averaged over the 24 distributed sources tested. The network also demonstrated robust learning capabilities after successfully training on noisy input data, with only relatively minor degradation to the source reconstruction quality. An analysis of variance study determined that the most significant factor affecting the reconstruction quality was the source's shape, with ring-type source distributions having the worst performance. The interaction between the source's size and direction was also discovered to have a small effect as larger sources located near the bottom of the system's field-of-view contained more phantoms within the reconstruction. Reconstruction quality was lower for responses exceeding the training noise level and for source distributions not included in the training set, indicating the importance of robust training data. The results show a significant improvement over more conventional algorithms, suggesting that real-time gamma imaging with the rotating scatter mask may be not only plausible, but practical for the first time. ReGeNN may readily be adapted for similar time-encoded radiation imaging systems, but the neural network methods described also have significant application potential towards other imaging systems.