We propose a novel cerebellum segmentation method for MRI, based on a combination of statistical models of the structure's expected location in the brain and its local appearance. The appearance model is obtained from a k-nearest-neighbor classifier, which uses a set of multi-scale local image descriptors as features. The spatial model is constructed by registering multiple manually annotated datasets to the unlabeled target image. The two components are then combined in a Bayesian framework. The method is quantitatively validated in a leave-one-out experiment using 18 MR images of elderly subjects. The experiment showed that the method produces accurate segmentations. The mean Dice similarity index compared to the manual reference was 0.953 for left and right, and the mean surface distance was 0.49 mm for left and 0.50 mm for right. The combined atlas- and appearance-based method was found to be more accurate than a method based on atlas-registration alone.