A robust and accurate method is presented for the segmentation of the cranial cavity in computed tomography (CT) and CT perfusion (CTP) images. The method consists of multi-atlas registration with label fusion followed by a geodesic active contour levelset refinement of the segmentation. Pre-registration atlas selection based on differences in anterior skull anatomy reduces computation time whilst optimising performance. The method was evaluated on a large clinical dataset of 573 acute stroke and trauma patients that received a CT or CTP in our hospital in the period February 2015 to December 2015. The database covers a large spectrum of the anatomical and pathological variations that is typically observed in everyday clinical practice. Three orthogonal slices were randomly selected per patient and manually annotated, resulting in 1659 reference annotations. Segmentations were initially visually inspected for the entire study cohort to assess failures. A total of 20 failures were reported. Quantitative evaluation in comparison to the reference dataset showed a mean Dice coefficient of 98.36 +- 2.59%. The results demonstrate that the method closely approaches the high performance of expert manual annotation.