Real-time adaptive proton therapy: An AI-based spatio-temporal mono-energetic dose calculation model (CC-LSTM)

Abstract

Purpose: To develop a fully AI-based dose estimation model capable of learning and estimating single pencil beam dose distributions, and to verify its performance by testing the model's generalizability on unseen, previously delivered treatment plans. Additionally, the model aims to achieve super-fast runtimes suitable for incorporation into real-time adaptive proton therapy (APT). Methods: A mono-energetic, end-to-end PB dose estimation task was defined using input Relative Stopping Power (RSP) and corresponding output dose distributions. A cohort of 90 Low-Grade-Glioma (LGG) patients was used for training and testing. The proposed CC-LSTM model employs 2-layer CNNs to extract spatial features from Beam's Eye View (BEV) slices, followed by a custom ConvLSTM to propagate 2D features along the beam path. A 3-layer CNN then reconstructs 2D dose distributions, which, in an auto-regressive scheme, form the 3D dose distribution of a single PB. Results: CC-LSTM demonstrated notable accuracy improvements over the RNN-based model, with the average local gamma-index pass rate at [1 %, 2 mm] increasing from 92.54% to 97.21% and the worst-case minimum rising from 71.69% to 92.37%, underscoring the robustness of the proposed AI-based dose estimation model. Additionally, CC-LSTM outperformed the current state-of-the-art (SOTA) model, achieving a notable decrease of up to three orders of magnitude in the MSE, faster runtimes, and a 98.8% reduction in the number of learnable parameters compared to the SOTA model. Conclusions: CC-LSTM can effectively learn the dose estimation task and generalize to unseen patient cases, achieving accuracies comparable to the gold-standard Monte Carlo simulations for highly heterogeneous cases, while maintaining runtimes suitable for incorporation into real-time APT.