ESTRO 2022 - Abstract Book

S10

Abstract book

ESTRO 2022

SP-0036 EFOMP: Next-generation CT: Innovations from photon-counting CT

M. Kachelrieß

Germany Abstract not available

Proffered Papers: Intra-fraction and real-time motion management

OC-0038 Sub-second speed proton dose calculation with Monte Carlo accuracy using deep learning

O. Pastor-Serrano 1 , Z. Perkó 1

1 Delft University of Technology, Radiation Science and Technology, Delft, The Netherlands

Purpose or Objective Radiotherapy workflow heavily relies on calculating the spatial distribution of physical dose within patients, among others for treatment planning and plan robustness assessment purposes. While ideally this calculation is quick and precise, current analytical pencil beam algorithms and stochastic Monte Carlo (MC) dose calculation tools offer a trade-off between accuracy and computational cost. Recently proposed deep learning (DL) methods attempt to solve this dichotomy by offering high speed and accuracy, although focusing on specific treatment plans/sites or using noisy/low accuracy dose distributions as input, which limit their generalization to other clinical settings or applications. To boost calculation times and ultimately make real-time adaptive treatments possible, we present a generic sub-second speed dose calculation algorithm accurately predicting the dose deposited by mono-energetic proton pencil beams for arbitrary energies and patient geometries. Materials and Methods For training and testing, we generate a dataset consisting of pairs of (i) 180x32x32 sized input CT fragments (with 2mm resolution in each dimension) containing relative stopping power values and (ii) the output dose distribution delivered by proton beamlets obtained from MC simulations with 10 billion source particles in 5 head and neck, 5 lung and 5 prostate patients. Framing proton transport as modeling a sequence of 2D geometries in the beam eye’s view as the particles travel through the patient, our dose algorithm processes the 3D stopping power images as a sequence of 2D slices and is trained to predict the ground-truth MC dose distributions. We combine convolutional neural networks extracting spatial features (e.g., tissue and density contrasts) with the Transformer self-attention mechanism that routes information between the elements in the sequence (i.e., different parts of the volume) and a vector representing the beam’s energy. Results Using a test dataset with patients unseen during training, we compare the model’s predictions to ground-truth MC simulations via gamma analysis Γ (1%, 3mm). With an average gamma pass rate of 99.6±0.76% and an absolute error always below 1.16% of the maximum dose, the model achieves close to MC accuracy even in the most heterogeneous geometries. Compared to computationally demanding MC simulations, our approach results in much faster calculation times, with an average of 34.05±0.5 ms per pencil beam (vs. the ~50 s for MCsquare calculations).