3D-2D Image Registration
How it works
The parameters of a 3D transformation (usually rigid: 3 translations + 3 rotations) are defined. In the intensity-based method, for a given volume pose a DRR (simulated X-ray projection) is generated, compared with the real 2D image using a similarity measure (e.g., mutual information, gradient difference), and an optimizer iteratively refines the pose until the similarity is maximized. In the feature-based method, projected 3D features are matched to features detected in the 2D image. The result is a transformation that overlays the 3D volume onto the 2D image in real time.
Problem solved
Preoperative 3D imaging (CT/MR) is rich in detail but unavailable during the procedure, where the operator has only 2D projections (fluoroscopy). Without a common coordinate system, the 3D plan and structures cannot be mapped onto the live image.
Evolution
Adoption of intensity-based 3D-2D registration using digitally reconstructed radiographs (DRRs) to match CT to X-ray.
Systematic comparison of similarity measures (incl. mutual information) for CT-to-fluoroscopy registration established best practices for the intensity-based method.
Markelj et al. systematized 3D/2D registration methods for image-guided interventions, becoming a field reference.
Deep-learning methods began estimating pose and accelerating 3D-2D registration, improving robustness to local minima.