A multi-atlas approach is proposed for the automatic segmentation of nine different structures in a set of head and neck CT images for radiotherapy. The approach takes advantage of a training dataset of 25 images to build average head and neck atlases of high-quality. By registering patient images with the atlases at the global level, structures of interest are aligned approximately in space, which allowed multi-atlas-based segmentations and correlation-based label fusion to be performed at the local level in the following steps. Qualitative and quantitative evaluations are performed on a set of 15 testing images. As shown by the results, mandible, brainstem and parotid glands are segmented accurately (mean volume DSC>0.8). The segmentation accuracy for the optic nerves is also improved over previously reported results (mean DSC above 0.61 compared with 0.52 for previous results).