Towards a fast and accurate simulation framework for 3D spherical source localization in the near field of a coded aperture gamma camera

Handheld gamma cameras with coded aperture collimators are under investigation for intraoperative imaging in nuclear medicine. Unlike other types of collimators, a coded aperture offers the possibility of 3D imaging of small spherical sources in the near field. However, due to the non-linear system model this option has rarely been investigated. We hypothesize that a deep learning approach is capable of 3D localization of spherical sources with an accuracy that is required for sentinel lymph node biopsy. In order to generate a sufficient amount of training data and densely sample the field of view Monte Carlo simulations are computationally too expensive. Thus, a fast and accurate simulation framework called ConvSim is presented. It computes the detector image from a 3D volume distribution of gamma sources in front of the gamma camera described as a voxel cube. By slice-wise convolutions with the corresponding point spread functions (PSF) a simulation time of few seconds is reached. For spherical sources non-linear near field effects are considered additionally. Comparing the generated detector images of exemplary spherical sources at different distances to Monte Carlo simulations yield a good correspondence with a multi-scale structural similarity index (MS-SSIM) between 0.65 and 0.91. Even though ConvSim entirely ignores the photon energy, the simulation framework serves as a fast and accurate alternative to time-consuming Monte Carlo simulations. For the future, the framework might be useful for investigating new types of coded apertures or the reconstruction of extended sources. ConvSim is available at https://github.com/tomeiss/convsim.