In medical diagnosis, high resolution (HR) images are indispensable for giving more correct decision. The super resolution technique, which can generate HR images from LR images based on machine learning, attracts hot attention recently. However, the conventional learning based SR generally cannot recover high frequency information. In this paper, we integrate a further learning step into the conventional method, and proposes a two-step learning based SR, which is prospected to recover most high frequency information lost in the available LR input. Furthermore, we also propose to use HR axial plane images of input volumes as HR training data to reconstruct HR coronal plane and sagittal plane images.