Purpose: A method that could enable dose calculations to be performed using magnetic resonance (MR) images for conventional treatment planning and adaptive planning using MR‐accelerator systems would be to apply bulk electron densities to the MR images. However currently bone must be manually segmented making this impractical. This work develops and tests an atlas‐based method to automatically segment bone on pelvic MR images for dose calculations. Methods: An MR whole‐pelvic atlas was created using manually delineated scans from 39 patients. Atlas‐based pelvic‐bone auto‐segmentations were then created for 25 patient scans using deformable image registration of the atlas to each patients scan with a leave‐one‐out atlas approach. These and corresponding expert manual segmentations were compared using the Dice similarity coefficient. Bone was assigned a density of 1.19 g/cm3 and all other tissues a water equivalent density. Treatment plans were generated on the whole‐pelvis MR images and doses compared for the manual and auto‐segmented bone plans. Results: The average Dice coefficient was 0.83 (standard deviation = 0.05). The average manual bone volume was 834.6 cm3 compared to the atlas based average volume of 840.0 cm3, with a mean difference of 5.4 cm3 (0.64%). The average ICRU point dose calculated on the MR images using the atlas‐based bone segmentations was 0.2% lower (standard deviation 0.3%) than the dose calculated using the manual bone segmentations. Conclusions: The atlas based method for auto‐segmentation of pelvic bone enables MR‐based prostate radiotherapy dose calculations for treatment and adaptive planning. © 2011, American Association of Physicists in Medicine. All rights reserved.