Deep learning-based cone beam CT correction for adaptive proton therapy

Cone beam computed tomography (CBCT) is an imaging modality frequently used in radiotherapy for daily patient alignment. Besides the use for patient positioning, CBCT images can also provide valuable information about changes of the patient anatomy. However, due to the relatively poor image quality of CBCTs compared to conventional computed tomography, the useability of CBCTs in radiotherapy is currently limited. Especially adaptive proton therapy workflows, which aim to adapt treatment plans to changes in patient anatomy, could greatly benefit from daily information about changes in patient anatomy seen on CBCTs. Therefore, this thesis focuses on (deep learning) approaches to improve CBCT image quality and to enable daily proton dose calculations.

The thesis investigates various CBCT correction techniques and evaluates their proton dose calculations accuracy in head and neck, and lung cancer patients. For the lung, a dynamic 4D-scenario, accounting for respiratory motion, was also investigated. In both anatomical regions, the results showed that deep learning can correct CBCTs and enable accurate proton dose calculations. Furthermore, a deep learning-based correction seems promising for future implementation in adaptive proton therapy workflows. The thesis also addresses the need for robust quality control procedures of corrected CBCTs by investigating a patient specific quality control procedure based on proton radiography.