Preview

PDF, English - main document Download (36MB) | Terms of use

Abstract

Biomechanical patient modeling incorporates physical knowledge of the human anatomy into the image processing that is required for tracking anatomical deformations during adaptive radiation therapy, especially particle therapy. In contrast to standard image registration, this enforces bio-fidelic image transformation. In this thesis, the potential of a kinematic skeleton model and soft tissue motion propagation are investigated for crucial image analysis steps in adaptive radiation therapy. The first application is the integration of the kinematic model in a deformable image registration process (KinematicDIR). For monomodal CT scan pairs, the median target registration error based on skeleton landmarks, is smaller than (1.6 ± 0.2) mm. In addition, the successful transferability of this concept to otherwise challenging multimodal registration between CT and CBCT as well as CT and MRI scan pairs is shown to result in median target registration error in the order of 2 mm. This meets the accuracy requirement for adaptive radiation therapy and is especially interesting for MR-guided approaches. Another aspect, emerging in radiotherapy, is the utilization of deep-learning-based organ segmentation. As radiotherapy-specific labeled data is scarce, the training of such methods relies heavily on augmentation techniques. In this work, the generation of synthetically but realistically deformed scans used as Bionic Augmentation in the training phase improved the predicted segmentations by up to 15% in the Dice similarity coefficient, depending on the training strategy. Finally, it is shown that the biomechanical model can be built-up from automatic segmentations without deterioration of the KinematicDIR application. This is essential for use in a clinical workflow.