Speaker
Description
The characterization of Bremsstrahlung spectra generated by electron accelerators is becoming increasingly crucial, particularly in radiation processing applications such as sterilization of medical devices or food irradiation. The growing transition from isotopic to electric irradiators presents new challenges related to the control of beam properties. In this context, the technologies and resources centre Aerial is looking to develop a tool and methodology in order to characterize, on-line, the properties of the Bremsstrahlung spectra produced in the installation feerix®. This step is very important for their irradiation operations to ensure precise dose deposition in the sample and to prevent activation processes by exceeding the photonuclear reaction threshold of the photons generated in the installation. However, conventional direct and indirect methods are limited in meeting the constraints of on-line measurement of Bremsstrahlung spectra produced by electron accelerators. In previous work we carried out on this topic, we presented a method based on the measurement of a dose distribution and the use of inverse methods to reconstruct the Bremsstrahlung spectra from the dose distribution. For this purpose, we introduced a simple and compact experimental setup that enables on-line measurement of the dose distribution. This setup consists of a scintillator irradiated at its edge and a sCMOS camera positioned above the scintillator to capture the resulting scintillation photons. From the 2D distribution representing the number of photons collected by the camera, we are then able to retrieve an equivalent of the dose distribution in the scintillator along the beam axes. In this initial approach, we then used conventional regularization methods (Tikhonov) to reconstruct the Bremsstrahlung spectra from the dose distribution. In this study, we propose a new theoretical approach using Neural Networks to address the ill-posed inverse problem, using the same experimental setup. This approach is motivated by the limitations of previously discussed regularization methods, where the regularization parameter is highly sensitive to the shape of the incident spectra and to the noise level in the measured dose distribution. To accurately reconstruct the Bremsstrahlung spectra, the regularization parameter would need to be optimized for each reconstruction, which is incompatible with on-line reconstruction. Furthermore, regularization methods are limited in their ability to estimate the maximum energy of the reconstructed spectra, a critical parameter for Aerial. In this context, our new approach appears to be more efficient and better suited for on-line reconstruction, provided that our training dataset is representative of the problem. Thus, we focus here on an analytical approach for generating realistic training and validation datasets, consisting of Bremsstrahlung spectra and their corresponding dose distributions in the scintillator. The architecture of the proposed neural network is also presented, along with a multi-parameter study focused on optimizing the experimental setup. Specifically, we evaluate the method’s performance with different scintillators, leading to varied dose distribution profiles, and with varying pixel sizes in the measured dose distribution. The neural network approach will be compared to regularization methods, and recommendations will be made for developing an experimental setup that maximizes reconstruction efficiency according to this new method.
