3D Reconstruction of Interventional Material from Very Few X-Ray Projections for Interventional Image Guidance

Today, minimally invasive endovascular interventions are usually guided by 2D fluoroscopy, i.e. a live 2D X-ray image. However, 3D fluoroscopy, i.e. a live 3D image reconstructed from a stream of 2D X-ray images, could improve spatial awareness. 3D fluoroscopy is, however, not used today, since no appropriate 3D reconstruction algorithm is known. Existing algorithms for the real-time reconstruction of interventional material (guidewires, stents, catheters, etc.) are either only capable of reconstructing a single guidewire or catheter, or use too many X-ray images and therefore too much dose per 3D reconstruction. The goal of this thesis was to reconstruct complex arrangements of interventional material from as few X-ray images as possible. To this end, a previously proposed algorithm for the reconstruction of interventional material from four X-ray images was adapted. Five key improvements allowed to reduce the number of X-ray images per 3D reconstruction from four to two: a) use of temporal information in a rotating imaging setup, b) separate reconstruction of different types of interventional material enabled by the computation of semantic interventional material extraction images, c) compensation of stent motion by spatial transformer networks, d) per-projection backprojection and e) binarization of the guidewire extraction images. While previously only single curves could be reconstructed from two newly acquired X-ray images, the proposed pipeline can reconstruct stents and even stent-guidewire combinations. Submillimeter reconstruction accuracy was demonstrated on measured X-ray images of interventional material inside an anthropomorphic phantom with simulated respiratory motion. Measurements of the dose area product rate of the proposed 3D reconstruction pipeline indicate a dose burden roughly similar to that of 2D fluoroscopy