Medical imaging quality is limited by physical, technology and economical considerations. In the practical application, bilinear or bicubic interpolation are performed to enhance the resolution of medical image. However, these interpolation methods will lead to undesirable jagged artifacts and blur details, which would affect the performance of post-processing such as segmentation and registration. Example learning based single image super-resolution (SR) is an economical way to recover the high-resolution (HR) images from a low-resolution (LR) image. The learning based methods are to add the high frequency information borrowed from the learning data to the input image by referencing the similarities between the input data and the learning examples. Due to the incompatibility between the examples and the testing image, this paper proposes a novel multiscale rotational similarity based single image SR method, in which the learning examples are generated from the LR image itself. The motivation of the proposed method is that the small patches in medical image tends to repeat themselves across different scale or different angle. With the intensive investigation of the similarity redundancy in the multiscale and multiangular levels, the relationship between the patches of HR image and that of LR images will be estimated. Subsequently, the HR images will be generated from a single LR image in a gradual magnification framework. Meanwhile, a nonlocal prior regularization terms will be utilized to enhance the accuracy of reconstruction. Some experiments are performed to indicate the feasibility and effectiveness of the proposed algorithm.