A Framework for Semi-automatic Fiducial Localization in Volumetric Images

Fiducial localization in volumetric images is a common task performed by image-guided navigation and augmented reality systems. These systems often rely on fiducials for image-space to physical-space registration, or as easily identifiable structures for registration validation purposes. Automated methods for fiducial localization in volumetric images are available. Unfortunately, these methods are not generalizable as they explicitly utilize strong a priori knowledge, such as fiducial intensity values in CT, or known spatial configurations as part of the algorithm. Thus, manual localization has remained the most general approach, readily applicable across fiducial types and imaging modalities. The main drawbacks of manual localization are the variability and accuracy errors associated with visual localization. We describe a semi-automatic fiducial localization approach that combines the strengths of the human operator and an underlying computational system. The operator identifies the rough location of the fiducial, and the computational system accurately localizes it via intensity based registration, using the mutual information similarity measure. This approach is generic, implicitly accommodating for all fiducial types and imaging modalities. The framework was evaluated using five fiducial types and three imaging modalities. We obtained a maximal localization accuracy error of 0.35 mm, with a maximal precision variability of 0.5 mm.