The document presents a method for predicting computed tomography (CT) images from magnetic resonance imaging (MRI) data using learned nonlinear descriptors and k-nearest neighbor (KNN) regression techniques. It addresses challenges in PET/MR imaging and radiation therapy by employing local sparse correspondence and diffeomorphic mapping to improve accuracy in CT estimation from MRI. The proposed method was evaluated on a dataset of 13 subjects with MRI and CT pairs to validate its effectiveness in predicting pseudo-CT images.