We present a novel algorithm for Computed Tomography (CT) that simultaneously computes a reconstruction and a corresponding segmentation. Our algorithm uses learned dictionaries for both the reconstruction and the segmentation, constructed via discriminative dictionary learning using a set of corresponding images and segmentations. We give a detailed description of the implementation of our algorithm, and computer simulations demonstrate that our method provides better results than the other SRS or dictionary-based methods, especially when there are not sufficient projections. Moreover, due to the regularization, the segmentations from our method has more smooth class interfaces.