Sparsity seeking total generalized variation for undersampled tomographic reconstruction

Here we present a novel iterative approach for tomographic image reconstruction which improves image quality for undersampled and limited view projection measurements. Recently, the Total Generalized Variation (TGV) penalty has been proposed to establish a desirable balance between smooth and piecewise-constant solutions. Piecewise-smooth reconstructions are particularly important for biomedical applications, where the image surface slowly varies. The TGV penalty convexly combines the first and higher order derivatives, which means that for some regions (e.g. uniform background) it can be more challenging to find a sparser solution due to the weight of the higher order term. Therefore we propose a simple heuristic modification over the Chambolle-Pock reconstruction scheme for TGV which consists of adding the wavelet thresholding step which helps to suppress aliasing artifacts and noise while preserve piecewise-smooth appearance. Preliminary numerical results with two piecewise-smooth phantoms show strong improvement of the proposed method over TGV and TV penalties. The resulting images are smooth with sharp edges and fewer artifacts visible.