Existing medical image registration algorithms generally have problems such as manual intervention and long registration time. In this paper, we propose a novel method based on superpixelwise principal component analysis (SuperPCA) and block-matching for registration between magnetic resonance imaging (MRI) and computed tomography (CT) brain images. Firstly, we apply SuperPCA to align MR and CT images roughly. We use SuperPCA to compute initial parameters which can prevent from falling into a local optimum. Next, registration is refined via a free-form deformation which is based on an block-matching algorithm. Here we adopt robust self-similarity descriptor as a similarity metric in both stages of registration, which can extract modality-invariant neighbourhood representations separately. Experimental results have shown that the performances of the obtained images are improved by comparing with the state-of-the-art methods, demonstrating potentials of the proposed method to be applied for brain tumor radiation therapy.