Intensity modulated proton therapy (IMPT) is a promising modality for the treatment of head and neck cancers, where target volumes are often surrounded by critical organs at risk (OAR). However, the sharp dose fall-off created by proton beams makes IMPT more sensitive than photon radiation therapy against setup variations and anatomical changes, which are common amongst head and neck patients. Over the course of a fractionated treatment, these effects can lead to a severe degradation of the treatment plan which can seriously compromise the benefits of IMPT over photon radiation therapy. In this presentation, we investigate the potential of adaptive proton therapy using cone-beam CT (CBCT) data in order to restore the original treatment plan quality online, at each fraction. First, we introduce a deep learning-based scatter correction method enabling real-time dose calculation on CBCT images. Second, we retrospectively evaluate the impact of online daily adaptive IMPT by comparing this approach to two state-of-the-art robust optimization methods on a representative cohort of 10 patients. Using fast CBCT scatter correction, deformable image registration and GPU accelerated Monte Carlo dose calculation, our in-house developed adaptive workflow achieved robust and accurate plan adaptation in clinically acceptable time. Within our patient cohort, the daily adapted plans yielded equivalent to superior target coverage with respect to both robust optimization methods, while significantly reducing the dose to most OARs as well as the integral dose to healthy tissues. Globally, our results suggest that online adaptive proton therapy is clinically feasible and has the potential to substantially improve the quality of head and neck IMPT treatments.