ESTRO 2024 - Abstract Book

S3676

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

ESTRO 2024

Figure 2: Gamma pass rates (3%/3mm, 2%/2mm, 1%/1mm) for the treatment plans (A-C) and the individual segments (D-F) of the test data set.

Conclusion:

In this work, a deep learning dose modelling framework was trained and tested. Model performance was tested on a clinical, highly heterogenous dataset in the MR-linac setting. Using a 3%/3mm gamma criterion, high median acceptance rates for both, individual segments and treatment plans were achieved, comparable to experimental quality assurance procedures. The dose distributions could be derived in few second, compared to several hours of computation time for the MC-simulation. Few individual segments showed low agreement with the MC-simulation, even for a 3%/3mm criterion potentially due to highly complex scenarios such as material boundaries or heterogeneous tissue densities. To take this into account, uncertainty modelling seems to be required in a next step before using the network in clinical routine, for example as a secondary dose calculation.

Keywords: deep-learning, MR-Linac

References:

[1] Friedel M et al. Med Phys 2019.

2509

Mini-Oral