To estimate the internal structures of the patient brain, several methods map a brain atlas to a patient brain shape using some landmarks, which are selected manually from the brain shape. The determination of the correspondence between small sulci of arbitrary two brains is complex. Moreover, the relationship between the surface shape and internal structure of the brain is unclear. Therefore, even if the surface the deformed atlas is fitted to the patient brain shape, the accurate internal structures of the brain can not always be obtained. To solve these problems, this paper proposes a method of selecting landmarks to estimate the internal structures of the patient brain by deforming brain atlas. Firstly, the brain shapes are represented by a simple shape. Secondly, a small number of initial landmarks are selected manually from the contours of the approximated brain shape and the internal structures which can be identified from MR images. Thirdly, some new landmarks are generated automatically on the contours based on the initial landmarks. Finally, the brain atlas is deformed non-rigidly using these landmarks. To evaluate our method, we estimate the patient brain structure using 4 methods. From the experimental results using 10 brain images, our method can estimate the patient brain structure reliably and stably using a small number of landmarks compared with the method using the landmarks on the original brain shape.