ESTRO 2024 - Abstract Book

S4614

Physics - Optimisation, algorithms and applications for ion beam treatment planning

ESTR0 2024

937

Mini-Oral

Prediction of the dose-averaged proton linear energy transfer using convolutional neural networks

Aaron Kieslich 1,2 , Sebastian Starke 2,3 , Martina Palkowitsch 1,2 , Fabian Hennings 1,2 , Esther G. C. Troost 1,2 , Mechthild Krause 1,2 , Jona Bensberg 4 , Christian Hahn 1,4,5 , Feline Heinzelmann 6,7 , Christian Bäumer 4,6 , Armin Lühr 4 , Beate Timmermann 6,7 , Steffen Löck 1,8 1 OncoRay—National Center for Radiation Research in Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, Technische Universität Dresden, Helmholtz-Zentrum Dresden–Rossendorf, Dresden, Germany. 2 Helmholtz-Zentrum Dresden - Rossendorf, Institute of Radiooncology – OncoRay, Dresden, Germany. 3 Helmholtz Zentrum Dresden - Rossendorf, Department of Information Services and Computing, Dresden, Germany. 4 TU Dortmund University, Department of Physics, Dortmund, Germany. 5 RaySearch Laboratories AB, Service department, Stockholm, Sweden. 6 West German Proton Therapy Centre Essen (WPE), University Hospital Essen, Essen, Germany. 7 Clinic for Particle Therapy, University Hospital Essen, Essen, Germany. 8 Department of Radiotherapy and Radiation Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, Technische Universität Dresden, Dresden, Germany Proton therapy is characterised by its conformal dose distribution, providing a higher level of normal tissue sparing compared to photon therapy. While the relative biological effectiveness (RBE) is currently assumed to be 1.1 in clinical practice, research indicates that the RBE of protons can vary based on factors such as the linear energy transfer (LET) [1]. This variability can lead to an increase in biological effectiveness at the distal end of the beam that is currently not fully considered in treatment planning. An unconsidered increase in RBE may lead to an increased occurrence of radiation-induced side-effects [2]. The current standard for calculating the LET is the utilization of Monte Carlo (MC) simulations. However, MC methods are computationally demanding and require a detailed model of the beam delivery system, a requirement often unattainable when analysing data from external centres or retrospective data. This constrains research on the variable RBE of protons, as the number of usable datasets is reduced. A deep learning (DL) model may offer a more efficient and universally applicable alternative to provide the LET. Therefore, this study investigates the potential of 3D convolutional neural networks (CNN) to approximate dose-averaged proton LET (LETd) distributions generated by MC simulations, using the planned dose distribution as input. Purpose/Objective:

Material/Methods:

The data used for training the CNN consists of 115 patients with primary brain tumours treated with pencil-beam scanning (PBS) proton therapy at the University Proton Therapy Dresden. To avoid the high uncertainty of the MC based LETd distributions in low dose regions, LETd distributions were clipped to a region of relevant dose with >2 Gy(RBE). The performance of three DL architectures (Unet, SegResNet and UNETR [3,4,5]) was evaluated using 5-fold internal cross-validation on the training cohort. The best architecture was then used for training a model on the full