ESTRO 2024 - Abstract Book

S4667

Physics - Optimisation, algorithms and applications for ion beam treatment planning

ESTR0 2024

A database with several experimental points extracted from the literature is also available so the user can easily benchmark the simulations’ results. It is also possible to import a different set of experimental points if needed.

To offer an alternative way to make predictions, a machine learning (ML) model was incorporated into the framework. It directly estimates cell survival according to cell features and irradiation conditions. The ML model, based on the Random Forest, was trained using data from the Particle Irradiation Data Ensemble (PIDE [7]), which compiles results from in-vitro experiments. The model was trained using a reduced dataset consisting of 923 lines. Experiments that reported less than 3 points were removed. For the training phase, we randomly selected 65% of the experiments while the remaining experiments were set aside for testing. To potentially predict deviations from the LQ model, columns related to alpha and beta were excluded. Each line was then split based on the number of measurement points, and the corresponding survival and dose data were added.

The model takes as input 11 variables, either continuous or discrete, that consider both physical and biological parameters. The R 2 obtained between prediction and ground truth was 0,83.

Results:

The created tool, named CELL (Cellular Effect and Lesions Linker, https://github.com/mpvalen/interfaz_grafica_adn) is a freely available graphical interface that enables the user to compute cell survival accounting for cell characteristics and beam properties (fig. 1). CELL has already proven to be able to reproduce survival data along the Spread-Out Bragg Peak (SOBP) for both proton [8-9] and carbon ions. Similar results can be achieved with the ML model. ML has the potential to predict experimental trends even in complex configurations. In Figure 2, the ML algorithm's predictions are shown, using spectra extracted by FLUKA in a SOBP configuration. While this may not serve as definitive proof of agreement, it is evident that the model successfully predicts the experimental trend.

Fig1: CELL interface