Feature-based depth estimation for monoplane X-ray imaging with limited C-arm motion

During interventional procedures, surgical guidance is provided through live X-ray imaging. Depth cues from 2D imaging can enhance the insight to the true position of important structures. In this paper, we examine a novel application where a small motion of the C- arm is used to create disparity between views for defining a sparse depth map of generic interest points. Feature points are detected and then matched across multiple views, and the depth of the points is estimated using multi-view geometry. Specific adaptations for our case are (1) the matching using geometric constraints for locally dense searching and (2) outlier rejection for robust feature tracks in multiple views. Evaluation using phantom images gave an accuracy in the mm-range for up to 20 views (or _ 300), and a sufficiently rich set of robustly tracked features. This creates a valid option for depth estimation of static points of clinical interest.