Modeling real world system geometry and detector response within a high-throughput x-ray simulation framework

Unique opportunities to sample parameter space using physics-based synthetic x-ray data

“…Recent work in synthetic data generation offers the potential to dramatically reduce the amount and types of empirical data collection required, while simultaneously enabling more rapid and successful algorithm development as well as system testing and evaluation [1,2]. There are, however, different types of synthetic data, including methods that add representative examples of items or scenarios based on training with empirical data (e.g., using a generative adversarial network, or GAN [3]), methods that numerically augment empirical measurements (e.g., data augmentation techniques [4]), and fully-virtual methods (e.g., physics-based synthetic data [5][6][7][8]). It is worth noting that only physics-based synthetic data, in which the virtual X-ray scanner and bags exist only as numerical representations, does not rely on an empirical instantiation of a real scanner or bag to produce data.…”

“…While we have previously shown that synthetic datasets have the capacity to directly match the statistics of an empirical dataset [6,7,9,10], approaching synthetic data validation and application from this perspective fundamentally limits the value of the synthetic data. In this paper, we discuss the ways in which synthetic datasets differ from empirical datasets and how these differences can result in more robust, interpretable detection algorithms and quantitative, complete assessments of a system's detection performance.…”