Content area
Head and neck cancer (HNC) is one of the most common cancers worldwide, characterized by a high incidence and persistently low survival rates, with approximately 50% of patients surviving beyond five years. Among the available treatment modalities, radiotherapy (RT) has emerged as a cornerstone due to its efficacy, lower invasiveness, and suitability for cases where tumors may be unresectable. However, despite continuous advancements in RT techniques, treatment-related toxicities remain a major challenge, significantly impacting patients' quality of life. Among these toxicities, dysphagia (difficulty in swallowing) and xerostomia (dry mouth) are among the most debilitating side effects, often leading to nutritional deficiencies, social limitations, and long-term complications. Improving toxicity prediction and mitigation requires three key advancements: (1) accurate segmentation of organs at risk (OARs) to ensure precise dose delivery, (2) refined methodologies for toxicity assessment and (3) dose-response modeling. This thesis addresses these aspects by developing deep learning (DL)-based segmentation frameworks, introducing an automated approach for dysphagia evaluation, and performing voxel-based analyses to identify xerostomia-related subregions. The first step toward improving toxicity prediction is ensuring accurate OAR segmentation. We developed and validated a deep learning framework for the automatic segmentation of 25 OARs in HN CT images, training with partially labeled datasets and longitudinal data. Our model demonstrated state-of-the-art performance, achieving Dice Similarity Coefficients exceeding 70% and Average Surface Distances below 2 mm. A clinical evaluation by an external radiation oncologist showed that 71% of automatically generated contours were acceptable without editing, surpassing the quality of manually drawn contours. Additionally, we introduced an uncertainty-based classification method to detect segmentation outliers, further enhancing reliability for clinical implementation. Building upon these advancements, we extended our segmentation model to cone-beam computed tomography (CBCT) to facilitate adaptive radiotherapy (ART). Given CBCT’s lower image quality and lack of well-annotated datasets, we build upon a synthetic CT-based segmentation pipeline, achieving comparable accuracy to CT-based segmentations. Dosimetric validation confirmed that CBCT-based contours maintained consistency with planned dose distributions, supporting the integration of automatic segmentation into ART workflows. With accurate OAR segmentation models established for both CT and CBCT images, we explored novel methods for toxicity assessment and prediction. Dysphagia lacks standardized and objective evaluation criteria. To address this, we developed an automated analysis framework for videofluoroscopic swallowing studies. This approach integrated DL-based region detection, dynamic parameter extraction, and machine learning, enabling a quantitative assessment of swallowing function. This is the most comprehensive automated VFSS analysis to date, setting the foundation for objective dysphagia characterization and future predictive modeling in HNC patients. For xerostomia, several objective assessment methods are available, including stimulated salivary flow. However, traditional xerostomia prediction models rely on global dose metrics, overlooking local dose-response relationships. To refine these models, we conducted a voxel-based analysis, identifying two subregions—within the contralateral parotid and tubarial glands—that exhibited stronger associations with xerostomia than conventional dose metrics. Their clinical relevance was validated using a functional PSMA-PET-based salivary activity template, reinforcing the importance of considering local dose effects in treatment planning. To account for anatomical changes during treatment, we extended our voxel-based analysis to accumulated dose maps, generated through deformable image registration. This analysis confirmed the tubarial gland’s critical role in xerostomia development, demonstrating that radiation dose to this structure strongly correlates with toxicity risk. Importantly, our findings indicate that while ART reduces mean parotid dose, it may inadvertently increase dose to the tubarial glands, potentially increasing the probability of xerostomia. These results underscore the need to incorporate accumulated dose information and voxel-based dose constraints into RT planning for improved toxicity mitigation. Overall, this thesis advances segmentation techniques, enhances toxicity assessment methodologies, and introduces refined dose-response models to improve toxicity prediction in HNC RT. Together, these contributions pave the way for more personalized, toxicity-aware radiotherapy strategies to improve long-term outcomes for HNC patients.