Session Item

The objective of this work is to propose a method to predict electron beam characteristics of a linear accelerator Monte Carlo model using machine learning, thus speeding up and automating the modelling process. The method is developed and tested based on a MLC-equipped stereotactic radiotherapy accelerator model.

The M6 Cyberknife (Accuray, Sunnyvale CA) equipped with the Incise2 MLC is first modelled in BEAMnrc/DOSXYZnrc using water phantom measurements. The model is then integrated in our home-developed Monte Carlo (MC) 3D dose verification platform Moderato, and validated on patient plans. The general principle of the machine learning (ML) algorithm is described in Figure 1. To generate the data necessary to train the algorithm, a series of simulated dose profiles and percentage depth doses (PDD) are created from the MC model for the 115x100mm MLC field, while varying electron beam spot size (from 1 to 4 mm) and energy (from 4 to 8 MeV). A Ridge regression algorithm is first trained to predict energy and spot size by splitting the simulated curves into training and test sets, allowing to finetune the parameters of the algorithm. The ML model is then applied on measured data provided by five other M6-equipped institutions, and the predicted values are introduced in a MC model to generate simulated profiles for the 115x100mm MLC and 5 mm fixed beams, to be compared with the measurements.



Figure 1. Principle of the electron beam characteristics prediction.

For our M6 device, optimal agreement between simulated and measured profiles in the water tank was reached for a monoenergetic electron beam of 6.75 MeV with a gaussian spatial distribution of 2.4 mm FWHM. Re-calculation of patient plans showed a good agreement (<2%) between the TPS algorithms and Moderato. During the optimization of the ML algorithm, cross-validation showed that electron beam energy and spot size could be predicted with a mean absolute error (MAE) of 0.1 MeV and 0.3 mm respectively. Predicted values for the five other M6 systems varied significantly from one device to another (Table 1). Simulations generated from these predicted values demonstrated a close agreement (<3%) between simulated and measured dose curves, except for centre #1 where differences went up to 6% (figure 2), which might be due to an asymmetry in the measured dose profiles for that specific machine.


Table 1. Predicted electron beam spot sizes and energies for the five M6 devices.

A machine learning algorithm was developed and tested to validate the concept of predicting electron beam parameters from profile data. This method would allow significant time gain in the context of integrating multiple accelerator models in a Monte Carlo-based dose verification system.