Given that the traditional method of similarity weighted fusion in the process of multi-atlas based medical image segmentation does not consider the interference and redundancy of the atlas set, a method of brain magnetic resonance (MR) image segmentation based on a two-stage atlas selection strategy is proposed. In this method, a method based on minimum angle regression is used for rough atlas selection. Then, a method based on the Hausdorff distance is adopted for target-oriented precise atlas selection. The rough selection method can find the atlas similar to the target image on the whole, remove invalid variables, and reduce the interference and redundancy of the atlas set. The precise selection method pays more attention to the similarity calculation of the target tissue, and the similarity results are not affected by the size and location of the target tissue. Experimental results show that the proposed method is more robust and accurate than the traditional one-stage atlas selection method based on similarity calculation of the rectangular region.