Enhanced difference ghost shuffle U-Net: A hybrid Monte Carlo deep learning denoiser for boron neutron capture...

Boron neutron capture therapy (BNCT) is a form of binary radiotherapy based on the 10B(n, α)7Li capture reaction. Current BNCT treatment planning system (TPS) relies on Monte Carlo (MC) simulations for dose calculation, which is time-consuming to achieving clinically acceptable statistical uncertainty. Recent deep learning (DL) approaches accelerate MC dose calculation but often require excessive computational resources and exhibit significant dose deviations in high-dose gradient regions, limiting clinical applicability.

We propose the EDGS-UNet, a hybrid MC-DL denoising framework, which can rapidly and automatically perform denoising on the 3D BNCT treatment dose distribution, enabling fast and accurate dose calculation.

EDGS-UNet performs denoising from two perspectives: enhancing voxel resolution and reducing the number of particles. Specifically, noisy BNCT dose components (D_B (boron dose), D_N (nitrogen dose), D_H (hydrogen dose), and D_γ (gamma dose)) are generated by MC simulation with a small number of particles in a coarse-grid phantom. Then they are refined to a fine-grid noise-free dose by a lightweight 3D U-Net. This architecture incorporates: (1) Enhanced Differential Ghost Convolution (EDGConv) for high-frequency gradient preservation without requiring additional parameters or computational overhead, (2) A Hierarchical Channel-Spatial-Pixel Attention (HCSPA) mechanism for adaptive feature refinement, and (3) Physically component-specific energy-conservation loss functions to constrain the denoising process. The model is trained and evaluated on a glioblastoma dataset (56 train/9 val/11 test cases) using TOPAS MC simulations.

Statistical evaluations-including mean square error (MSE), isodose similarity, mean absolute percentage error (MAPE), and relative deviation of GTV-were performed. These results demonstrate that EDGS-UNet outperforms state-of-the-art architectures (3D U-Net, Attention U-Net, and GhostU-Net) in dose denoising, particularly in accurately preserving high-gradient regions. For the dose to 98% and 95% of the GTV volume (D98 and D95), the relative deviation between the denoising results and the simulation results is less than 3% and 3.5% respectively. This value is much lower than that of other architectures. The MAPE values of all organs computed by EDGS-UNet are approximately 2%-5%, less than 3D-UNet (6%-12%), Attention-UNet (5%-11%), and Ghost-UNet (3%-13%). Furthermore, the parameters used by EDGS-UNet are only one-tenth of those used by other competing models. The dose volume histogram generated by EDGS-UNet is highly consistent with the results obtained through MC simulation, and the generated dose map has the smallest error from the ground truth dose map, further confirming its accuracy.

EDGS-UNet offers a plug-and-play, clinically promising solution for rapid and accurate BNCT dose calculation. This work is expected to promote the clinical development of BNCT in the future.