A GPU-accelerated framework for individualized estimation of organ doses in digital tomosynthesis

Purpose: Estimation of organ doses in digital tomosynthesis (DT) is challenging due to the lack of existing tools that accurately and flexibly model protocol- and view-specific collimations and motion trajectories of the source and detector for a variety of exam protocols, and the computational inefficiencies of conducting MC simulations. The purpose of this study was to overcome these limitations by developing and benchmarking a GPU-accelerated MC simulation framework compatible with patient-specific computational phantoms for individualized estimation of organ doses in DT. Materials and methods: The framework for individualized estimation of dose in DT was developed as a two-step workflow: (1) a custom MATLAB code that accepts a patient-specific computational phantom and exam description (organ markers for defining the extremities of the anatomical region of interest, tube voltage, source-to-image distance, angular sweep range, number of projection views, and the pivot point to image distance - PPID) to compute the field of views (FOVs) for a clinical DT system, and (2) a MC tool (developed using MC-GPU) modeling the configuration of a clinical DT system to estimate organ doses based on the computed FOVs. Using this framework, we estimated organ doses for 28 radiosensitive organs in an adult reference patient model (M; 30 years) imaged using a commercial DT system (VolumeRad, GE Healthcare, Waukesha, WI). The estimates were benchmarked against values from a comparable organ dose estimation framework (reference dataset developed by the Advanced Laboratory for Radiation Dosimetry Studies at University of Florida) for a posterior–anterior chest exam. The resulting differences were quantified as percent relative errors and analyzed to identify any potential sources of bias and uncertainties. The timing performance (run duration in seconds) of the framework was also quantified for the same simulation to gauge the feasibility of the workflow for time-constrained clinical applications. Results: The organ dose estimates from the developed framework showed a close agreement with the reference dataset, with percent relative errors ranging from −6.9% to 5.0% and a mean absolute percent difference of 1.7% over all radiosensitive organs, with the exception of testes and eye lens, for which the percent relative errors were higher at −18.9% and −27.6%, respectively, due to their relative positioning outside the primary irradiation field, leading to fewer photons depositing energy and consequently higher errors in estimated organ doses. The run duration for the same simulation was 916.3 s, representing a substantial improvement in performance over existing nonparallelized MC tools. Conclusions: This study successfully developed and benchmarked a GPU-accelerated framework compatible with patient-specific anthropomorphic computational phantoms for accurate individualized estimation of organ doses in DT. By enabling patient-specific estimation of organ doses, this framework can aid clinicians and researchers by providing them with tools essential for tracking the radiation burden to patients for dose monitoring purposes and identifying the trends and relationships in organ doses for a patient population to optimize existing and develop new exam protocols.