Fast CBCT Reconstruction using Convolutional Neural Networks for Arbitrary Robotic C-arm Orbits.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Radiology

Background:

• Non-circular orbits in Cone-Beam CT (CBCT) offer improved image quality and field-of-view for interventional imaging.
• Rapid reconstruction is crucial for time-sensitive interventional procedures.
• Model-Based Iterative Reconstruction (MBIR) is computationally intensive for arbitrary geometries.

Purpose of the Study:

• To develop and evaluate a fast reconstruction framework for non-circular CBCT trajectories.
• To reduce the computational burden of CBCT image reconstruction while maintaining image quality.
• To enable clinically relevant reconstruction times for advanced CBCT acquisition methods.

Main Methods:

• A pipeline combining deconvolution with an approximate system response and a Convolutional Neural Network (CNN) was developed.
• The CNN was trained on 1800 randomized arbitrary orbits using simulated data from phantoms and real patient images (LIDC).
• Reconstruction performance was evaluated using quantitative metrics (nRMSE, FSIM, SSIM) and compared against MBIR.

Main Results:

• The proposed pipeline achieved a 90% reduction in computation time compared to MBIR with minimal performance differences.
• Reconstruction performance remained consistent even for acquisition orbits not encountered during CNN training.
• Quantitative metrics demonstrated the efficacy of the fast reconstruction method.

Conclusions:

• The developed framework enables fast processing of arbitrary CBCT trajectory data.
• Reconstruction times are clinically relevant, facilitating the use of non-circular orbits in image-guided interventions and intraoperative imaging.
• This approach holds potential for advancing interventional CT imaging applications.