In medical image practice, it is common to upsample images for better view or subsequent processing, such as image segmentation or registration. The commonly used upsampling methods include bicubic or cubic splines. The shape of interpolation kernels are usually square windows. However, the square kernels are not always optimal since there are a lot of complex local structures in medical images. The interpolated high-resolution pixels should preserve these local structures. In this paper we propose a novel shape-adaptive method for magnetic resonance (MR) images upsampling. First, the optimal ...