Robust neural network-based unfolding of bremsstrahlung spectra from depth dose measurements in radiotherapy

In radiotherapy, quality control of medical linear accelerators mainly relies on absolute dose and depth dose measurements in water phantoms, while the bremsstrahlung spectrum is not routinely monitored. However, accurate knowledge of this spectrum is essential for reliable Monte Carlo simulations and advanced dose calculations used to complement treatment planning systems (TPS). Direct spectral measurements are challenging due to the pulsed beam structure and high dose rates, which can saturate or damage conventional spectrometers. Several indirect techniques have been proposed, including transmission measurements, Compton scattering and photoactivation methods, but they often require complex or invasive experimental setups. An approach based on unfolding the bremsstrahlung spectrum from depth dose distributions has also been explored in this context. Despite proposed regularization strategies to mitigate the ill-posed nature of the problem, these methods remain highly sensitive to noise and exhibit limited accuracy in reproducing physically realistic spectra. In this work, we propose a novel reconstruction approach to unfold bremsstrahlung spectra from depth-dose measurements in a water phantom. The method relies on a parametric representation of the spectrum, enabling the generation of physically realistic datasets. A neural network is trained to infer the seven parameters defining this representation, enforcing strong physical constraints on the spectral shape. The approach was validated using experimental measurements acquired with a 6 MV Varian TrueBeam accelerator at the Institut de Cancérologie Strasbourg Europe (ICANS). Compared with previously published methods, the proposed approach demonstrates improved efficiency, enhanced spectral realism, and increased robustness to noise in the dose distribution.