Linear Object Registration for Image-Guided Interventions

Abstract

Purpose: Image-guided interventions rely on registration of images, models, and surgical tools into a common navigation space. Point-set registration is commonly used to perform this registration, but requires well-defined landmark points to be present on these tools. In this thesis, a generalization of point-set registration is proposed to simultaneously register point, line, and plane landmarks present on surgical tools. This facilitates registration when point-set registration is not feasible. Methods: The proposed algorithm first determines correspondences between points, lines, and planes in the coordinate systems using a set of “reference” landmarks, then calculates invariant features in each coordinate frame for an initial registration, and finally optimizes the registration iteratively. Several forms of validation are investigated: registration of simulated data with a known ground-truth registration, phantom registration using a tracked stylus for registration to a model or a volume, and volume registration of a reconstructed ultrasound volume to a model. Validation accuracy is determined by comparison to a known ground-truth registration or using registration quality metrics such as target registration error. Results: For the simulated data experiments, the linear object registration was sufficiently close to the ground-truth registration in all cases given the level of noise. For real registration experiments, in all instances where accurate point-set registration was possible, the linear object registration was equally as accurate, and the difference between the two registrations was less than the fiducial localization error. When accurate point-set registration was not possible, the linear object registration was observed to be more accurate and more precise than point-set registration using approximate landmarks. Conclusion: The proposed linear object registration algorithm is a viable alternative when point-set registration cannot be performed. The algorithm has been developed as an open-source registration tool for practical use as a module for the 3D Slicer platform.