A Study of Image-based C-arm Tracking Using Minimal Fiducials
Ayad, Maria Samir
MetadataShow full item record
Image-based tracking of the c-arm continues to be a critical and challenging problem for many clinical applications due to its widespread use in many computer-assisted procedures that rely upon its accuracy for further planning, registration, and reconstruction tasks. In this thesis, a variety of approaches are presented to improve current c-arm tracking methods and devices for intra-operative procedures. The first approach presents a novel two-dimensional fiducial comprising a set of coplanar conics and an improved single-image pose estimation algorithm that addresses segmentation errors using a mathematical equilibration approach. Simulation results show an improvement in the mean rotation and translation errors by factors of 4 and 1.75, respectively, as a result of using the proposed algorithm. Experiments using real data obtained by imaging a simple precisely machined model consisting of three coplanar ellipses retrieve pose estimates that are in good agreement with those obtained by a ground truth optical tracker. This two-dimensional fiducial can be easily placed under the patient allowing a wide field of view for the motion of the c-arm. The second approach employs learning-based techniques to two-view geometrical theories. A demonstrative algorithm is used to simultaneously tackle matching and segmentation issues of features segmented from pairs of acquired images. The corrected features can then be used to retrieve the epipolar geometry which can ultimately provide pose parameters using a one-dimensional fiducial. The problem of match refinement for epipolar geometry estimation is formulated in a reinforcement-learning framework. Experiments demonstrate the ability to both reject false matches and fix small localization errors in the segmentation of true noisy matches in a minimal number of steps. The third approach presents a feasibility study for an approach that entirely eliminates the use of tracking fiducials. It relies only on preoperative data to initialize a point-based model that is subsequently used to iteratively estimate the pose and the structure of the point-like intraoperative implant using three to six images simultaneously. This method is tested in the framework of prostate brachytherapy in which preoperative data including planned 3-D locations for a large number of point-like implants called seeds is usually available. Simultaneous pose estimation for the c-arm for each image and localization of the seeds is studied in a simulation environment. Results indicate mean reconstruction errors that are less than 1.2 mm for noisy plans of 84 seeds or fewer. These are attained when the 3D mean error introduced to the plan as a result of adding Gaussian noise is less than 3.2 mm.