ESTRO 2024 - Abstract Book

S3674

Physics - Dose prediction, optimisation and applications of photon and electron planning

ESTRO 2024

References:

1, ESTRO 2023 Pre-meeting course: A one-day SBRT/SRS bootcamp

2483

Proffered Paper

Deep learning based modelling of radiation dose deposition in a 1.5 T MR-Linac

Moritz Schneider 1 , Simon Gutwein 1 , Maximilian Nyazi 2,3 , Daniela Thorwarth 1,3

1 Section for Biomedical Physics, Department of Radiation Oncology, University Hospital Tübingen, Tübingen, Germany. 2 Department of Radiation Oncology, University Hospital Tübingen, Tübingen, Germany. 3 German Cancer Research Consortium (DKTK), Partner Site Tübingen, Tübingen, a partnership between DKFZ and University Hospital Tübingen, Germany

Purpose/Objective:

Monte-Carlo (MC) simulations are the current gold standard for dose calculation in radiotherapy. But the tradeoff between accuracy and computational effort prevents this technique from accurate dose calculation in real time. As the presence of a strong external magnetic field inhibits the use of fast analytical dose calculation algorithms, deep learning (DL) based dose modelling approaches promise drastic reduction of computation time while achieving accurate dose distributions. This would open new possibilities for online adaptive magnetic resonance (MR) guided adaptive radiotherapy. To evaluate the possibilities of this concept we developed a DL framework to train neural networks to model dose distributions, based on patient CT and radiation field information. Consequently, the aim of this study was to train a neural network on a comprehensive 1.5 T MR-Linac dataset and investigate its performance on unseen patient data against full MC simulations.

Material/Methods:

The training data was based on 3981 clinical radiotherapy treatment segments from 75 clinical patient plans treated at the 1.5 T MR-Linac (Unity, Elekta AB, Sweden), suffering either from prostate (40), mamma (5), head and-neck (10) or liver (10) cancer. Each data set consisted of the patient’s CT and RT-plan file defining every segment. As ground truth, we simulated the dose distribution for every patient and irradiation segment using the EGSnrc MC framework with a specific MR-linac model [1]. Using this data, a 3D-UNet was trained to model the dose distribution in 3D for given irradiation segments. We used a root mean squared error loss, the ADAM optimizer with a learning rate of 1E-4 and trained for 6000 Epochs, using a patch size of 32x32x32 voxel and a batch size of 128 patches. The trained model was then evaluated on a test data set consisting of 2657 segments of 50 treatment plans from 10 primary prostate patients, 10 patients suffering from head-and-neck tumors, 10 patients with tumors located in the liver, 5 patients with partial breast irradiation and 15 patients with cancerous lymph nodes. Note, that the last entity (lymph node) was not present in the training data. We used 3%/3mm, 2%/2mm and 1%/1mm gamma criteria with a 10% dose cutoff to evaluate the agreement of the DL based dose modelling to MC simulated data, again derived from the EGSnrc framework.

Results: