AI-driven anatomical patient modelling from PET data for enhanced adaptive charged particle therapy

Protons and stable carbon ions are highly effective at targeting radio-resistant and deep-seated tumours. However, there is a non-negligible uncertainty on the beam range within the patient, mostly due to changes in the patient’s anatomy. Non-invasive, in-vivo online treatment monitoring is essential. One of the ways is using in-beam positron emission tomography (IB-PET) scanner. However, there is a lack of a direct representation of the patient's anatomy in the IB-PET images. This thesis aims at addressing this issue by 1) evaluating the efficacy of analytical methods to extract inter-fractional anatomical changes information from the INSIDE IB-PET scanner data in both proton and 12C-ion beam therapy at CNAO. 2) Pioneering the application of deep learning techniques to create synthetic control CT images from combining IB-PET and planning CT scan data during proton therapy. 3) Exploiting IB-PET monitoring with radioactive ion beams. This thesis has resulted in three significant findings. The analytical methods have shown great capability of providing parameters that, if monitored, can provide indications of possible deviations during proton therapy treatment, monitored with the INSIDE IB-PET scanner. In stable carbon-ion beam therapy, no consistent results were found. AI models have demonstrated their ability to generate synthetic CT images effectively from inter-fractional activity maps data of proton-treated patients and planning CT scans but were less effective when using IB-PET data from a planar PET scanner. Employing radioactive ion beams can enhance the IB-PET imaging signal output of an order of magnitude compared to stable ion beams, and it opens the door for NN models to enhance RIBs treatment accuracy using IB-PET data.