Reconstruction of bone microstructure from few projections with convex-concave and non local regularization

This work investigates a new method based on discrete tomography for the reconstruction of binary cross-sections of bone micro-structure from a small number of projections. While priors based on Markov random fields have been previously considered for the reconstruction of binary images, we propose to improve the regularization term by introducing long range smoothness constraints. To this aim, we propose to use a convex-concave deterministic optimization approach coupled with a non local regularization. Applications to 256×256 bone cross-section images provides good results from only 20 projections, even in the presence of additive gaussian noise.