Heidelberg 2022 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

ST: Fachverband Strahlen- und Medizinphysik

ST 3: Artificial Intelligence in Medicine

ST 3.3: Vortrag

Dienstag, 22. März 2022, 11:30–11:45, ST-H4

Using generative adversarial networks to predict proton beam dose distributions in mice — •Lara Bußmann, Kevin Kröninger, Armin Lühr, Florian Mentzel, Janine Salewski, and Jens Weingarten — TU Dortmund

The clinically used generic relative biological effectiveness (RBE) of 1.1 for protons compared to photons does not consider variations along the beams axis. For a better estimation of the varying RBE and to assess potential adverse effects, mouse brains are irradiated and excised to visualize DNA double-strand breaks.

In order to deduct conclusions about the RBE, the observed irradiation damage in the tissue is compared to the expected damage from Monte Carlo simulations of the dose distribution.

Using Monte Carlo simulations for dose distribution predictions can be very time-consuming. Machine learning models can be trained to predict dose distributions based on the phantom geometry.

In this talk, a deep learning dose prediction model for proton mouse irradiations based on generative adversial networks (GANs) is presented. GANs can be trained to generate data samples following a learnt distribution, which are indistinguishable from a ground truth distribution. In this study, MC simulation samples are used to train the GAN, using geometrical information about the target phantom as conditional input.