We propose to develop and evaluate robust deep learning (DL)-based approaches capable of accurately delineating target volumes and predicting recurrence in head and neck cancer (HNC) patients. Radiation therapy (RT) is one of the most common treatments for HNC patients. Advanced RT techniques enable highly conformal dose delivery to target volumes. However, a major challenge in the RT planning for HNC is delineating target tumor volumes. Despite the availability of consensus guidelines, delineating the gross target volume (GTV) and the clinical target volume (CTV) for HNC is time-consuming and requires extensive clinical expertise. It demands a comprehensive understanding of the region's intricate anatomy, tumor histology, and spread patterns. Precise delineation of target volumes, especially the CTV, is essential to avoid marginal misses and excess doses to organs at risk (OARs). Also, existing DL target volume delineation algorithms have predominantly focused on the GTV delineation of oropharyngeal cancer, neglecting other commonly encountered tumor subsites, such as laryngeal and nasopharyngeal cancers, which represent ~30% of HNC. Another challenge in HNC management is the high recurrence rate. There is a daunting 30% five-year recurrence rate, with at least 50% occurring in-field, despite comprehensive treatment strategies encompassing surgery, chemotherapy, and RT. 18F-FDG-PET/CT has become part of the standard of care for HNC thanks to its ability to improve the accuracy of GTV delineation, reveal previously undiagnosed regional nodal disease, and contribute to a decrease in inter- and intra- observer variability (IOV) in GTV delineation. However, most DL algorithms have used contours derived from a single physician as the gold standard (label), failing to capture IOV in tumor delineation, an essential component of robust DL strategies. We will develop robust delineation algorithms capable of accurately delineating both GTV and CTV in various HNC locations in both primary and recurrent settings. We propose diffusion-based DL algorithms to delineate GTV and CTV from 18F-FDG-PET and contrast-enhanced CT, while capturing observer variability. We will also develop a DL method that incorporates imaging and clinical information to predict recurrence and whether recurrence will occur in-field.