Small Beams, Fast Predictions - Comparison of Modern Machine Learning Models for Microbeam Radiation Therapy
F Mentzel1*, M Barnes2, K Kröninger1, M Lerch2, O Nackenhorst1, J Paino2, A Rozenfeld2, A Tsoi3, J Weingarten1, M Hagenbuchner3, S Guatelli2, (1) Department of Physics, TU Dortmund University,DE, (2) Centre For Medical Radiation Physics, University Of Wollongong, AUS, (3) University Of Wollongong, AUS
Presentations
PO-GePV-M-37 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: Treatment planning for novel treatments such as synchrotron and proton microbeam radiation therapy is often done using full Monte Carlo simulations for dose prediction. Different deep learning algorithms have been proposed to accelerate this step. We evaluate the usability of selected deep learning models with focus on the small sizes which are important for microbeam radiation therapy and compare the models systematically in terms of accuracy and execution speed.
Methods: We compare two recently published models: the first is based on a 3D U-Net which can be trained as regression (using e. g. the mean squared error as cost function) or as generative adversarial network (GAN) using a second neural network to provide a dynamic cost function. The second model is derived from a novel transformer-based dose prediction approach that encodes CT slices into digital word tokens and “translates” the series of CT tokens into a series of Dose tokens which is then expanded into a dose prediction. The training data is obtained by means of Geant4 simulations using a simplified digital skull phantom.
Results: All models are significantly faster than non-machine learning approaches to dose prediction and provide reasonably accurate results in less than one second. Overall, we find the transformer-based model to be most flexible while the 3D U-Net trained with mean squared error as cost function provides the most accurate predictions both in the synchrotron and proton microbeam scenario. The GAN trained 3D U-Net is found to be less stable and accurate compared to the regression-trained model.
Conclusion: Small-scale beam therapies are getting increasingly noticed using both sub-millimetre gamma and proton beams. Deep learning models such as transformer-based and 3D U-Net models can significantly accelerate the dose prediction. As the models were compared using simple phantoms, the transferability into application is limited.
Funding Support, Disclosures, and Conflict of Interest: The authors gratefully acknowledge the computing time provided on the Linux HPC cluster at Technical University Dortmund (LiDO3), partially funded in the course of the Large-Scale Equipment Initiative by the German Research Foundation (DFG) as project 271512359.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email