Optimizing dosimetry in Y-90 microsphere radioembolization: GPU-accelerated Monte Carlo simulation versus...

Yttrium-90 (⁹⁰Y) microsphere radioembolization has shown unique advantages in treating both primary and metastatic liver cancer and was introduced into China in 2022. Despite the development of various dosimetric models-ranging from empirical to voxel-based approaches-practical implementation remains challenging. With over 370,000 new liver cancer cases annually and limited access to certified ⁹⁰Y treatment centers, Chinese interventional oncology departments face increasing pressure to balance dosimetric accuracy with clinical efficiency. This study aims to develop a GPU-based fast Monte Carlo project for accurate voxel-level dose calculation and to evaluate its performance alongside existing dosimetric strategies, with the goal of supporting optimized clinical workflows in high-volume settings.

A fast Monte Carlo simulation algorithm was developed using Graphics Processing Unit (GPU) acceleration and applied retrospectively to eight patients diagnosed with hepatocellular carcinoma or metastatic colorectal cancer. The dosimetric performance of the GPU-based approach was compared against direct Monte Carlo (MC) simulations, the Medical Internal Radiation Dose (MIRD) formalism, the Voxel S-value (VSV) method, and the Local Energy Deposition (LED) model. Voxel- and organ-level dose accuracy were quantified using metrics such as Mean Absolute Relative Error (MARE), Relative Standard Deviation (RSD), and D95 in dose volume histogram. Statistical comparisons were conducted using Shapiro-Wilk normality tests and repeated measures ANOVA to assess inter-method differences.

The GPU-based Monte Carlo code demonstrated high accuracy and computational efficiency. Using direct MC simulation as the reference, the GPU-based approach yielded the lowest voxel-level variability, with median RSDs in high-activity transverse regions reaching - 1.13%, indicating superior consistency. Corresponding MARE were 4.53% for the GPU method, compared to 6.71% for VSV and 49.36% for LED, confirming its dosimetric reliability. At the organ level, the GPU-based method achieved RSDs of 0.35% ± 0.80% (tumor), -0.45% ± 0.76% (liver), 1.41% ± 4.45% (lung), and - 1.43% ± 1.23% (spleen), significantly outperforming alternative models. Notably, VSV and LED substantially underestimated lung dose (-52.19% ± 23.87%, -53.71 ± 22.17%), highlighting their limited applicability in heterogeneous regions. In contrast, the dose of spleen (F = 3.26, p = 0.069) and kidneys (F = 3.22, p = 0.071) did not show statistically significant differences between methods. In terms of computational performance, the GPU-based code delivered a remarkable 1,296-fold speed-up over traditional MC simulations, enabling efficient voxel-level dosimetry suitable for clinical workflows.

The GPU-based fast Monte Carlo simulation provides a highly accurate and computationally efficient tool for voxel-level dosimetry in ⁹⁰Y radioembolization. It enables precise estimation of tumor and lung doses with significantly reduced processing time and hardware demands, offering clear clinical advantages in minimizing radiation pneumonitis risk and supporting high-throughput workflows. Importantly, a stratified approach to dosimetric modeling-selecting simplified methods such as VSV or LED for small, well-contained lesions, and reserving GPU-based Monte Carlo for anatomically complex or heterogeneous cases-may optimize the balance between accuracy and efficiency. Future work will focus on large-scale validation and formalizing model selection criteria tailored to tumor morphology and treatment scope, with the aim of advancing personalized dosimetric planning in liver-directed therapies.

Not applicable.