ESTRO 2025 - Abstract Book

S3436

Physics - Machine learning models and clinical applications

ESTRO 2025

complexity of DNA damage, remains a significant challenge. An example is the proton spread-out Bragg peaks (SOBP), where LET variations at a constant dose increase lethality, especially near the distal edge. This work integrates machine learning (ML) models and Monte Carlo (MC) simulations to predict cell survival fraction (SF) along the SOBP for in-vitro experiments with heavy particles. Material/Methods: A Random Forest (RF) model was trained on data extracted from the Particle Irradiation Data Ensemble (PIDE) [1] version 3.2, which recompiles 1118 in-vitro experiments from the literature. The dataset was preprocessed to filter experiments reporting less than three dose-survival points and negative dose or survival values. A set of 130 cell lines and 26 ions was used. The square of the dose (Dose2) was added as an additional parameter. A feature selection was performed and six features were chosen: Dose, Dose2, Energy, Charge and Cells. For complex distributions like an SOBP, the FLUKA [2] MC code was used. The SF is calculated using the ML model for each depth using the dose and energy fluences simulated with FLUKA as input. The MC simulation presented in this work has been validated against experimental data [3]. It's based on an experiment with protons, and the cell line used was Chinese hamster V79 [4]. Results: The RF strongly correlated with experimental survival data (Fig. 1), accurately predicting survival rates across diverse radiation types and energies with an R 2 = 0.806. In Fig. 2, the ML algorithm's predictions are shown in a SOBP configuration and it successfully predicts the experimental trend in the distal region.

Fig. 1: SF extracted from PIDE (testing set) plotted against the predictions by the ML algorithm.

Fig. 2: Survival versus depth following irradiation for a 230 MeV proton SOBP. Orange squares correspond to the ML predictions and blue “X” represent the experimental data from [4].