Prompt-gamma imaging (PGI) was proposed in the 2000s as a promising in vivo range-verification method to maintain the physical advantage of proton beams by reducing unwanted range-uncertainties. Recently, PGI with a slit camera has been successfully implemented in clinical application. Despite its high accuracy and sensitivity to range deviation being shown in several studies, the clinical benefits of PGI have not yet been investigated. Hence, to fully exploit the advantages of PGI, this thesis aims to investigate the feasibility of PGI-based range verification for the automatic classification of treatment deviations and differentiation of relevant from non-relevant changes in the treatment of head-and-neck (H&N) tumors. In the first part of this thesis, the four most common types of treatment deviations in proton therapy (PT) were investigated regarding their PGI signature and by considering clinically relevant and non-relevant scenarios. A heuristic decision tree (DT) model was iteratively developed. To gain understanding of the specific signature of the error sources, different levels of geometrical complexities were explored, from simple to complex. At the simplest level, a phantom with homogeneous density was used to distinguish range-prediction and setup errors. Next, in the intermediate complexity level, a phantom with heterogeneous density was used to inspect the additional error scenarios of anatomical changes. Finally, real patient CT scans were used to investigate the relevance of changes based on clinical constraints. In the final model, a five-step filtering approach was used during pre-processing to select reliable pencil-beam-scanning spots for range verification. In this study, five features extracted from the filtered PGI data were used to classify the treatment deviation. The model is able distinguish four introduced scenarios into six classes as follows: (1) overestimation of range prediction, (2) underestimation of range prediction, (3) setup error with larger air gap, (4) setup error with smaller air gap, (5) anatomical change, and (6) non-relevant change. To ensure the application was effective, independent patient CT datasets were used to test the model. The results yielded an excellent performance of the DT classifier, with high accuracy, sensitivity, and specificity of 96%, 100%, and 85.7%, respectively. According to these findings, this model can sensitively detect treatment deviations in PT based on simulated PGI data. In the second part of this work, an alternative approach based on machine learning (ML) was taken to automatically classify the error sources. In the first stage, the two approaches were compared, using the same features as well as the same training and test datasets. The results show that the ML approach was slightly better than the heuristic DT approach in terms of accuracy. However, the performance of both approaches was excellent for the individual scenarios. Thus, these results confirm that the PGI-based data classification with five features can be applied to detect individual sources of treatment deviation in PT. In the second stage, there was an investigation of more complex and more realistic combinations of error scenarios, which was out of the scope of the DT approach. The results demonstrated that the performance of the ML-based classifiers declined in general. Furthermore, the additional features of the PG shift did not substantially improve the performance of the classifiers. As a consequence, these findings mark important issues for future research. Potentially, usage of the spatial information from the spot-based PGI data and more complex techniques such as deep learning may improve the performance of classifiers with respect to scenarios with multiple error sources. However, regardless of this, it is recommended that these findings be confirmed and validated in simulations under measurement-like conditions or with real PG measurements of H&N patients themselves. Moreover, this classification model could eventually be tested with other body sites and entities in order to assess its compatibility and adaptation requirements. In summary, this study yielded promising results regarding the automatic classification of treatment-deviation sources and the differentiation of relevant and non-relevant changes in H&N-tumor treatment in PT with PGI data. This simulation study marks an important step towards fully automated PGI-based proton-range verification, which could contribute to closing the treatment-workflow loop of adaptive therapy by supporting clinical decision-making and, ultimately, improving clinical PT.:1 Introduction 2 Background 2.1 Proton therapy 2.1.1 Rationale for proton therapy 2.1.2 Uncertainties and their mitigation 2.2 In vivo range-verification techniques 2.2.1 Range probing 2.2.2 Proton tomography 2.2.3 Magnetic resonance imaging 2.2.4 Ionoacoustic detection 2.2.5 Treatment-activated positron-emission tomography imaging 2.2.6 Prompt-gamma based detection 3 Prompt-gamma imaging with a knife-edged slit camera 3.1 Current state-of-the-art 3.2 Prompt-gamma camera system 3.3 Data acquisition and analysis 4 Error-source classification using heuristic decision tree approach 4.1 Study design 4.1.1 Case selection 4.1.2 Investigated scenarios 4.1.3 Prompt-gamma simulation and range shift determination 4.2 Development of the model 4.2.1 First-generation model 4.2.2 Second-generation model 4.2.3 Third-generation model 4.3 Model testing 4.4 Discussion: decision-tree model 5 Error-source classification using a machine-learning approach 5.1 Machine learning for classification 5.1.1 Support-vector-machine algorithm 5.1.2 Ensemble algorithm – random forest 5.1.3 Logistic-regression algorithm 5.2 Study design 5.2.1 Case selection 5.2.2 Feature selection 5.3 Model generation 5.4 Model testing 5.5 Discussion 6 Summary/ Zusammenfassung Bibliography Appendix List of Figures List of Tables List of Abbreviations

L'hadronthérapie est une méthode de traitement du cancer qui emploie des ions (carbone ou proton) au lieu des rayons X. Les interactions entre le faisceau et le patient produisent des radiations secondaires. Il existe une corrélation entre la position d'émission de certaines de ces particules et la position du pic de Bragg. Parmi ces particules, des gamma-prompt sont produits par les fragments nucléaires excités et des travaux actuels ont pour but de concevoir des systèmes de tomographie par émission mono-photonique capable d'imager la position d'émission ces radiations en temps réel, avec une précision millimétrique, malgré le faible nombre de données acquises. Bien que ce ne soit pas actuellement possible, le but in fine est de surveiller le dépôt de dose. La caméra Compton est un des système TEMP qui a été proposé pour imager ce type de particules, car elle offre une meilleure résolution énergétique et la possibilité d'avoir une image 3D. Cependant, en pratique l'acquisition est affectée par le bruit provenant d'autres particules secondaires, et les algorithmes de reconstruction des images Compton sont plus compliqués et encore peu aboutis, mais sur une bonne voie de développement. Dans le cadre de cette thèse, nous avons développé une chaîne complète allant de la simulation de l'irradiation d'un fantôme par un faisceau de protons allant jusqu'à la reconstruction tomographique des images obtenues à partir de données acquises par la caméra Compton. Nous avons étudié différentes méthodes de reconstruction analytiques et itératives, et nous avons développé une méthode de reconstruction itérative capable de prendre en compte les incertitudes de mesure sur l'énergie. Enfin nous avons développé des méthodes pour la détection de la fin du parcours des distributions gamma-prompt reconstruites. / Hadrontherapy is a cancer treatment method which uses ions (proton or carbon) instead of X-rays. Interactions between the beam and the patient produce secondary radiation. It has been shown that there is a correlation between the emission position of some of these particles and the Bragg peak position. Among these particles, prompt-gamma are produced by excited nuclear fragments and current work aims to design SPECT systems able to image the emission position the radiation in real time, with a millimetric precision, despite the low data statistic. Although it is not currently possible, the goal is to monitor the deposited dose. The Compton camera is a SPECT system that proposed for imaging such particles, because it offers a good energy resolution and the possibility of a 3D imaging. However, in practice the acquisition is affected by noise from other secondary particles and the reconstruction algorithms are more complex and not totally completed, but the developments are well advanced. In this thesis, we developed a complete process from the simulation of irradiation of a phantom by a proton beam up to the tomographic reconstruction of images obtained from data acquired by the Compton camera. We studied different reconstruction methods (analytical and iterative), and we have developed an iterative method able to consider the measurement uncertainties on energy. Finally we developed methods to detect the end-of-range of the reconstructed prompt-gamma distributions.