Marker-based C-arm self-calibration with unknown calibration pattern.

Area of Science:

• Medical Imaging
• Computed Tomography
• Geometric Calibration

Background:

• Accurate tomographic reconstruction relies on precise acquisition geometry.
• Mobile C-arm CT scanners often exhibit poor geometry reproducibility, necessitating specific calibration.
• Existing self-calibration methods have limitations, requiring prior information or estimating few parameters, or using impractical calibration patterns.

Purpose of the Study:

• To introduce a novel three-step marker-based self-calibration method for C-arm CT.
• To enable self-calibration without prior knowledge of the calibration pattern's marker positions.
• To facilitate the use of arbitrary marker configurations in calibration patterns.

Main Methods:

• Marker detection in projections using YOLO (You Only Look Once) Convolutional Neural Network.
• Estimation of projected marker trajectories via a Linear Assignment Problem scheme with Kalman filters.
• Acquisition geometry determination using the Bundle-adjustment algorithm.

Main Results:

• Successful testing on simulated ICRP phantom images with 10 and 20 markers, yielding sub-millimeter backprojection errors.
• Qualitative evaluation on real human knee images (10 and 30 markers) demonstrated significant artifact reduction.
• Improved visibility of bone structures in clinical C-arm CT images.

Conclusions:

• The proposed marker-based self-calibration method shows promising results for clinical applications.
• This technique paves the way for patient-specific geometric self-calibrations.
• The method enhances image quality by reducing artifacts and improving structural visibility.