Robots Atlas>ROBOTS ATLAS
Computer Vision

3D-2D Image Registration

1995ActivePublished: 17 November 2026Updated: 17 November 2026Published
Key innovation
Linking the rich preoperative 3D volume to the sparse but live 2D projections, providing full-anatomy navigation without additional intraoperative 3D imaging.
Category
Computer Vision
Abstraction level
Technika
Operation level
InferenceApplication
Use cases
Navigation in endovascular and neurovascular surgeryOverlaying the plan on fluoroscopy during thrombectomyImage-guided spine surgery and orthopedicsRadiotherapy and patient positioningSurgical robotics and instrument tracking

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

Original paper · 2012 · Medical Image Analysis 16(3), 642-661 · Primoz Markelj
A review of 3D/2D registration methods for image-guided interventions
Primoz Markelj, Dejan Tomazevic, Bostjan Likar, Franjo Pernus
1995
DRR-based registration

Adoption of intensity-based 3D-2D registration using digitally reconstructed radiographs (DRRs) to match CT to X-ray.

1998
Comparison of similarity measures

Systematic comparison of similarity measures (incl. mutual information) for CT-to-fluoroscopy registration established best practices for the intensity-based method.

2012
Canonical review of 3D/2D methods
Inflection point

Markelj et al. systematized 3D/2D registration methods for image-guided interventions, becoming a field reference.

2016
Learning-based registration

Deep-learning methods began estimating pose and accelerating 3D-2D registration, improving robustness to local minima.