ESTRO 2025 - Abstract Book

S2692

Physics - Dose calculation algorithms

ESTRO 2025

2060

Digital Poster Angle dependent dose transformer algorithm for fast proton therapy dose calculations Mikołaj Arkadiusz Stryja, Zoltán Perkó, Danny Lathouwers Radiation Science & Technology, Delft University of Technology, Delft, Netherlands

Purpose/Objective: Adaptive pencil beam scanning proton therapy (PBSPT) requires a fast and accurate dose calculation. Currently available deep-learning (DL)-based dose engine approaches, as presented in [3] and [4], require a CT grid input that is perpendicular to the proton beamlet’s ray. This constraint necessitates interpolating the input CT grid for each beamlet angle. The objective of this study is to develop a deep learning-based dose engine capable of accurately predicting 3D dose distributions delivered to the patient by proton beamlets with specified initial energy and angles determined by two scanning magnets. Material/Methods: We present the General Dose Transformer Algorithm (GDoTA), a novel transformer-based framework for predicting 3D dose distributions within a given patient geometry. Our model’s key innovation is the effective encoding of proton beamlet directions utilising fast 3D flux projections. These projections are constructed based on the beamlet spot size, beamlet angles, and initial energy of the beamlet determining initial spot size. Incorporating the beamlet direction as an input enables precise modelling of angular dependencies in dose distribution and using a single interpolated CT per gantry angle as geometry input, significantly reducing input preparation time. The model was trained on a dataset of 12,000 unique CT snippets extracted from 60 distinct lung CT scans from [1] using AdamW optimizer. Training was limited to 50 epochs, with an early-stopping mechanism. The model was trained using a balancing mean squared error (MSE) between predicted and true dose distributions and their corresponding Integral-Depth-Dose (IDD) profiles. Ground truth dose distributions were generated using MCsquare, a high-fidelity Monte Carlo dose calculation algorithm [2].