Image registration and hybrid volume reconstruction of bone anatomy using a statistical shape atlas
Russell H. Taylor, Ofri Sadowsky · 2009
Recovery of three dimensional (3D) anatomical structures is a necessity in image-assisted surgery. Intra-operative x-ray images acquired with a C-arm are traditionally reconstructed in the surgeon's mind, based on his anatomical knowledge. This thesis studies an emerging method for x-ray based reconstruction, where a computational pre-operative anatomical model replaces the surgeon's knowledge. The model can be derived from a patient-specific CT scan or from a statistical database (an anatomical atlas). A registration (alignment) of the prior model and the observed x-rays recovers the 3D shape and radiodensity of the patient's anatomy. This knowledge is then combined with the x-rays to fill-in for missing intra-operative information and improve the quality of a CT-like cone-beam We name the process of registration and fusion between the x-ray and atlas modalities reconstruction. Potentially, it can bring affordable and mobile 3D medical imaging to environments and organizations that could not use it before. To accomplish this goal, the thesis presents the following research contributions. We developed an efficient visualization algorithm to create 2D projections of our atlas. The algorithm, used as a component in registration and subsequent reconstrution, is key for the practical use of the atlas. We present an assessment of the registration of atlas and x-rays. And we demonstrate how hybrid reconstruction reduces artifacts caused by missing or distorted information in conventional reconstruction, due to image truncation, limited scan trajectory, and errors in tracking the C-arm motion. Importantly, we present results of both simulation and real x-ray data. The latter required development of calibration and scan protocols with a ubiquitous mobile C-arm, which is a minor yet important research contribution. Among our results are a 100X speedup in the rendering of atlas images with the new algorithm, compared with CPU computation, sub-millimetric accuracy of rigid registration in both simulation and x-ray experiments, about 2 mm mean error is deformable registration, low errors in simulated hybrid reconstruction, and successful hybrid reconstruction from x-rays fused with atlas projections.