Radiotherapy is an important modality in treatment cancer. In this domain, proton beams have ballistic superiority against photon beams. Among the existing delivery techniques, pencilbeam scanning is one of the most modern ones and presents many advantages in terms of flexibility and dose conformation. Nevertheless, the use of pencil beams (PBs) still requires well adapted dosimetry tools for small irradiation fields quality assurance. To answer this issue, an innovative dosimetry system has been developed. It is based on a scintillating block of 10 × 10 × 10 cm3, a mirror and an ultra-fast camera recording the scintillation from two orthogonal points of view to reconstruct 3-dimensional dose maps [1].
The recorded images require some extensive preprocessing to correct hot pixels caused by scattered radiation directly hitting the camera sensor, optical artifacts[2] and the scintillation quenching [3].
3D reconstruction usually relies on numerous projections, not achievable in dosimetry in radiotherapy and many reconstruction algorithms have been developed for imaging purpose.
A reconstruction method based on only two projections and relying on the radial symmetry of PBs has been developed and showed good performances. Nevertheless, some steps of the preprocessing and the reconstruction algorithm are computing time consuming and the radial symmetry limits its application to PBs.
Different experimental setups and machine Learning (ML) methods will be studied in this PhD thesis to address these limitations.
In a first step, two different setups (the previous setup and one including an additional scintillator sheet) generating different scintillation projections will be simulated to study different reconstruction strategies.
These digital setups will be fed to a treatment planning system (TPS) to generate 3D and 2D dose distributions for the two different setups. In parallel, LET and quenching correction maps will be computed to convert the dose distributions into scintillation distributions.
A projection matrix corresponding to the studied setups will allow to convert the 3D and 2D scintillation distributions into projection images and optical artifacts and noise (characterized in a preliminary M2 internship) will be added create a realistic images database.
The collection of 3D dose distribution and corresponding computed images will constitute the database used to study and evaluate different ML strategies for 3D dose reconstruction. The implemented reconstruction methods will be evaluated with experimental measurements performed at the Cyclhad proton therapy center, compared to conventional reconstruction methods and confronted to other beams such as not circular photon beams.