A C-arm is an X-ray imaging device used for minimally invasive interventional radiology procedures. Most modern C-arm systems are capable of rotating around the patient while acquiring radiographic images, from which a 3D reconstruction can be performed. This technique is called C-arm computed tomography (C-arm CT) and is used in clinical routine to image static organs. However, its extension to imaging of the beating heart or the free-breathing thorax is still a challenging research problem. This thesis is focused on human cardiac C-arm CT. It proposes several new reconstruction methods and compares them to the current state or the art, both on a digital phantom and on real data acquired on several patients. The first method, ECG-gated Iterative FDK deconvolution, consists in filtering out the streak artifacts from an ECG-gated FDK reconstruction in an iterative deconvolution scheme. It performs better than existing deconvolution-based methods, but it is still insufficient for human cardiac C-arm CT. Two 3D reconstruction methods based on compressed sensing are proposed: total variation-regularized 3D reconstruction and wavelets-regularized 3D reconstruction. They are compared to the current state-of-the-art method, called prior image constrained compressed sensing (PICCS). They exhibit results that are similar to those of PICCS. Finally, two 3D+time reconstruction methods are presented. They have slightly different mathematical formulations but are based on the same principles: using a motion mask to restrict the movement to the area containing the heart and the aorta, and enforcing smoothness of the solution in both space and time. One of these methods outperforms PICCS by producing results that are sharper and more consistent throughout the cardiac cycle
