Model-based 3-D LV shape recovery in biplane x-ray angiography: A-priori information learned from CT

Recovering the 3-D LV shape from non-rotational bi-planar x-ray angiograms is a very challenging task. The inherently sparse and noisy data available for reconstruction and the ill-posed nature of the inverse problem necessitate the incorporation of a-priori information. To this end, a statistical shape model of the LV anatomy is learned from high-resolution multi-slice CT data. Reconstruction is based on a non-rigid 2-D/3-D registration technique. To fit pose and shape of the model to the x-ray images of the patient, simulated projections of the model are calculated and the difference between given and simulated projections is minimized. The presented approach is evaluated using simulated and in-vivo angiograms. For patients where both CT and angiograms are available, the reconstructed LV is compared to the true shape known from CT. The defined similarity metrics used for evaluation show a good correspondence between recovered and true shapes.