DIGITALLY RECONSTRUCTED WALL RADIOGRAPHS
Xianhe Ee · National University of Singapore · 2008
Computer-Aided Surgery (CAS) technology enables the use of computers to generate 3D virtual environments of body parts slated for operation. In these virtual environments, surgeons enjoy the benefits of visualization, pre-operative planning, simulation and realtime navigation. All these translate into better surgical treatments, reduced complications, improved patient safety and lower health-care costs. CAS applications, like model-based segmentation, real-time navigation, simulation and modeling of orthopedic implants, require an essential technique for medical image analysis which is the registration of an anatomical model to medical images. Existing registration approaches include geometry-based and intensity-based approaches which have some shortcomings. Intensity-based approach generates a lot of irrelevant details which can obscure relevant features. Hence, it is susceptible to getting trapped at local minima. It is also computationally expensive. On the other hand, geometry-based approach needs to extract features from the images. Feature extraction algorithms generally cannot distinguish between relevant and irrelevant features. Thus, it is difficult to automate this approach with accuracy and reliability. There are three critical components in accurate and robust model-based medical image registration: the model, the objective function and the optimization algorithm. This thesis shall focus on the modeling of 3D data for registration. It proposes a hybrid approach that can combine the strengths of the two registration approaches while mitigating their weaknesses. It uses a 3D wall model of an anatomical part such that the wall surfaces capture surface shape, the solid wall captures intensity information, and the interior is hollow. In a digitally reconstructed radiograph of the wall model (DRWR), high-contrast features similar to edges and contours are produced only by the wall. DRWR can be registered to an x-ray image using intensity-based approach without the need for feature extraction.