Fast personalized CT dose calculations with GPUMCD.

[PURPOSE] The continuous increase of population dose due to ever-rising Computed Tomography (CT) examinations has called for more personalized dose estimations in medical imaging - a far from trivial task. This study aims to demonstrate a GPU-enabled pipeline combining automatic segmentation with GPU Monte Carlo Dose (GPUMCD) simulations to provide patient-specific dose-to-organ CT dosimetry reports using existing patient CT images.

[METHODS] A dynamic representation of the CT imaging process was reproduced within GPUMCD using information in DICOM headers, complemented by in-house exposure measurements, and validated in homogeneous and anthropomorphic phantoms. A dose pipeline was implemented using GPUMCD and a pre-trained open-source nnU-net model (TotalSegmentator). Dose-to-organ dosimetry was obtained for images from a lung cancer screening program and stored in DICOM-compliant Structured Reports.

[RESULTS] GPUMCD calculated dose values were within 5.5% of measurements for all phantoms and investigated conditions. Utilizing one A100-SXM4-40GB GPU, the average pipeline runtime was 6 min and 06 s per CT study. The GPU-driven simulation and segmentation operation took 46% (2 min and 7 s) of the total runtime, and data processing (file reading, conversion, and writing) occupied the remaining 54%.

[CONCLUSION] This work demonstrates the ability to generate patient-specific three-dimensional dose distributions in CT within a few seconds and the subsequent feasibility of performing fully automated mass personalized dose-to-organ calculations. The pipeline ingests and produces DICOM-compliant data compatible with clinical and research environments, enabling routine imaging dosimetry and large-scale retroactive dosimetry studies.