GMMC-IR: a GPU-accelerated fast dose calculation algorithm for interventional radiology

Abstract Prolonged fluoroscopy during interventional radiology (IR) procedures exposes both operators and patients to potential radiation injury risks. Accurate assessment of organ absorbed doses for both operators and patients in complex clinical scenarios is critical. Traditional central processing unit based Monte Carlo (MC) simulations are often limited in routine clinical workflows due to substantial computational demands. This work describes the development and validation of GMMC-IR graphics processing unit (GPU mesh MC for IR), a GPU accelerated MC code specifically tailored for IR applications to achieve rapid dosimetry while maintaining high physical fidelity. To address GPU thread divergence, an event-based approach was employed to port Geant4 physics functions into NVIDIA CUDA kernels, incorporating optimisations such as static memory pooling, event-pool-based particle encoding, and dynamic scheduling. For geometry navigation, the hardware-accelerated ray-tracing capabilities of NVIDIA OptiX™ were utilised. GMMC-IR was benchmarked and validated against Geant4 in a homogeneous water phantom, the ICRP adult female mesh-type phantom mesh-type reference computational phantom (MRCP_AF), and typical IR scenarios. Geant4 simulations were executed using 100 threads on a server equipped with dual AMD EPYC 7763 64-core processors and 512 GB of RAM, while GMMC-IR was run on an NVIDIA RTX 4090 GPU with 24 GB of VRAM. For coupled photon–electron transport in the water phantom, differences in depth dose and lateral dose profiles were within 3%. In the MRCP_AF and typical IR scenarios, organ dose differences remained within 5%. Compared to the Geant4 baseline, GMMC-IR achieved speedup factors of approximately 90–855. This study provides a viable tool for the rapid assessment of dose distributions for both operators and patients in IR settings.