S2281
Physics - Machine learning and AI algorithms
ESTRO 2026
This project aims to develop a machine-learning, surrogate-based digital twin for FLASHKNiFE that can automatically optimize accelerator parameters to target beam configurations. A recent study highlights the potential role of instantaneous dose rate in the FLASH effect [2]. Therefore, the digital twin’s first task was to generate configurations with different instantaneous dose rates while keeping other beam parameters constant for in vitro and in vivo experiments. Material/Methods:
The digital twin combines a surrogate model, trained to predict the machine’s output, with an optimizer identifying configurations that achieve target beam properties. The surrogate, whose structure is shown in Figure 1, was trained on a dataset of 400 generated accelerator configurations paired with beam parameters measured using a flashDiamond detector [3]. Model inputs consist of accelerator control parameters, while outputs correspond to dose metrics. The configuration set was designed to span the parameter space. The regression methods used include Neural Networks, Gaussian Processes, Random Forests, and XGBoost. Hyperparameters were optimized via Bayesian search, and performance evaluated using the Mean Absolute Error (MAE). The optimization followed a cascade architecture, in which an algorithm (Genetic Algorithm, Bayesian Optimization, or Convolution Matrix Approach– Evolutionary Strategy (CMA-ES)) iteratively generated candidate configurations while the surrogate evaluated each proposal. Results:
Conclusion: Anatomical augmentation resulted in large increases in dose prediction accuracy. When using anatomical augmentation, the number of real patients (R-patients) could be reduced by a factor of 5. Keywords: Anatomical augmentations, automated planning Digital Poster Highlight 575 Developing a machine learning-based digital twin for an electron FLASH radiotherapy LINAC Johan Pierre LEYGONIE 1,2 , Thibault Dijoud 2 , Aashini Rajpal 2 , Elena Agostoni 2 , Eric Deutsch 1 , Philippe Liger 2 , Charlotte Robert 1 1 UMR 1030 Molecular Radiotherapy and Therapeutic Innovation, Gustave Roussy, Villejuif, France. 2 Innovation & technologie, Theryq, Rousset, France Purpose/Objective: FLASH is an emerging radiotherapy (RT) technique that widens the therapeutic window compared to conventional dose rate RT. Given the novelty of this technology, research experiments to better understand the FLASH effect are needed. FLASHKNiFE is a medical linear accelerator (LINAC) delivering low- energy electron beams. Currently in preclinical use, it shows potential for ultra-high dose-rate radiotherapy [1]. One challenge is the manual tuning of the LINAC parameters, which is time-consuming and suboptimal.
The surrogate model can predict beam parameters from a given configuration, with the best normalized MAE and R-squared on the test set ranging from 5.40 to 6.38 % and 0.81 to 0.84, respectively (Table 1). The fitness score obtained with the optimizer during configuration generation is on the order of 10-2 to 10- 4 with CMA-ES, the most efficient method, for a calculation time of ~10 minutes per configuration. For the instantaneous dose-rate configurations generated
Made with FlippingBook - Share PDF online