Session Item

Fast dose computations are important for radiotherapy treatment planning, especially for plan optimization, requiring many variations before finalization. While fast approximations exist for most clinical treatments, they are not available for some specialized or novel treatments. One such novel treatment is microbeam radiation therapy (MRT), a pre-clinical technique which relies on arrays of sub-mm synchrotron-generated, polarized X-ray beams. MRT has been shown to exhibit improved healthy tissue sparing qualities.
Precise dose computation using Monte Carlo (MC) simulations is both time consuming and memory intensive due to the high resolution required to capture the dose gradients at the edge of the microbeam peaks. We investigate a time efficient alternative to full MC using generative adversarial networks (GANs) which are trained to accurately predict dose distributions for variable phantoms and irradiation scenarios.

The presented machine learning (ML) model comprises of a conditional 3D-UNet GAN, which learns to generate a dose deposition prediction based on a 3D CT scan. As proof of concept, we predict the dose depositions obtained using Geant4 for a broad synchrotron beam in a water phantom containing a bone slab of variable rotation angles and thicknesses. Subsequently, we demonstrate that our model is generalisable by applying it to a simplified head phantom MC simulation. Finally, we explore the transition to spatially highly confined microbeams for which we conduct a systematic characterization of field size effects using MC simulation of individual beams. To mitigate the memory limitations of the MC simulation, a novel data collection approach is introduced. 

For the broad synchrotron beam, the trained model predicts for both the bone slab inside the water phantom and the simple head phantom dose distributions with deviations of less than 1% of the maximum dose for over 94% of the simulated voxels in the beam. Dose predictions near material interfaces are accurate on a voxel-by-voxel basis with less than 5% deviation in most cases. Dose predictions can be produced in less than a second on a desktop PC compared to approximately 50 CPU hours needed for the corresponding Geant4 simulation.
The predicted peak and valley doses from arrays of microbeams using the novel MC data collection approach match previous MC simulations and are found to be suitable for the use as machine learning training data.

The presented ML model can be trained on Geant4 simulation data to generate accurate dose predictions in our experiments consisting of a bone slab in water and a simple head phantom in the case of a MRT broad beam. A systematic MC study on dose depositions from arrays of planar microbeams suggests that the model can be extended for MRT dose prediction as well. In future studies we want to include the model within a treatment planning system for MRT. The presented approach can likely be adapted for other novel treatment methods as well.