An AI dose-influence matrix engine for robust pencil beam scanning protons therapy

Background: Rapid planning is of tremendous value in proton pencil beam scanning (PBS) therapy in overcoming range uncertainty. However, the dose calculation of the dose influence matrix (D_ij) in robust PBS plan optimization is time-consuming and requires substantial acceleration to enhance efficiency.

Purpose: To accelerate the D_ij calculations in PBS therapy, we developed an AI-D_ij engine integrated into our in-house treatment planning system (TPS).

Methods: The AI-D_ij engine calculates spot dose using a transformer-based spot dose calculation model (SDM), which takes CT volumes (CT-bars, 256 $\times$ 16 $\times$ 16 voxels, 3 mm resolution) and energy (a float value) as inputs and outputs the spot dose distribution (256 $\times$ 16 $\times$ 16). The SDM was trained on over 200 000 CT-bars and Monte Carlo (MC) spot dose (spanning energy levels from 70 to 225 MeV). Clinical-implemented treatment plans for the head, lung, and liver, initially created on Raystation, were replanned using our AI-D_ij engine under identical gantry angles and uncertainties settings. After optimizing the spot weight, each in-house plan was recalculated using MCsquare for MC dose evaluation. The dose-volume histogram (DVH) metrics from the in-house TPS and Raystation were compared, evaluating both the optimized and MC doses.

Results: In optimization, the differences of DVH metrics (%, Value_in-house-Value_Raystation) across all uncertainty scenarios between the in-house and Raystation plans were 0.93 ± 2.04% for clinical target volume (CTV) and -5.94 ± 12.19% for organ at risks (OARs). For the MC doses, the differences were 2.48 ± 2.78% for CTV and -5.47 ± 14.16% for OARs. The time cost of a robust AI-D_ij calculation can be within 2s on an RTX3090 GPU.

Conclusion: We conducted a feasibility study on AI-D_ij engine-based robust PBS plan optimization, demonstrating both high planning speed and quality.