In interventional radiology, 3D volume data can be overlaid onto the real-time 2D X-ray projections (fluoroscopy) to augment with more spatial information.

Two issues that are mostly complained by the clinicians are accuracy and visualization.

The quality (or reliability) of 2D/3D overlay relies on the accuracy and the robustness of the registration algorithms. As part of pre-operative procedure, an initial 2D/3D registration is normally done at the beginning of the procedure (based on calibration information, optimization or even manual adjustment). However, patient movements can lead to inaccuracies in the overlay, because in many procedures no general anesthesia is applied. A rigid 2D/3D-registration is typically used for the estimation and correction of the patient movement.

Robust registration is challenging in clinical practice due to lots of reasons, such as difference of data dimension (3D and 2D), different possible imaging modalities (2D: fluoroscopy, US... 3D: X-ray CT, MR...), different possible motions from patient and external devices.

The key problems to be solved are:

• to recognize external devices or automatically ignore them
• to archieve 2D+t registration: motion prediction based on previous frames
• to make full use of 3D information