Abstract: Positron emission tomography computed tomography （PET/CT） is increasingly being used in modern radiotherapy （RT） treatment in combination with computed tomography （CT）. Due to the high dose of radiation exposure in CT scan，there is always a great desire to reduce the amount of the radiation dose in x-ray computed tomography （CT） because of the health risks. In this paper，we investigate the potential of non-local self-similar patch-based dictionary learning based pseudo CT estimating method for tackling this challenging task. Firstly，the method partitions training CT and MRI images into a set of patches. For each patch，voxel multi-scale features are extracted and a block-matching method is applied to cluster the CT image patches. Then by dictionary learning，we obtain the MRI-CT mapping matrix and predict CT patches as a structured output. Finally，by reconstruction method，we output the estimated CT image. Experimental results show that our method perform better than existing atlas-based method and showed a promising potential for RT of brain based only on MRI.