DRACO: differentiable reconstruction for arbitrary CBCT orbits.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Healthcare

Background:

• Traditional iterative algorithms for cone beam computed tomography (CBCT) reconstruction face significant computational and memory challenges.
• Reconstructing images from arbitrary or non-continuous X-ray trajectories is particularly demanding.

Purpose of the Study:

• To introduce a novel, efficient, and accurate method for CBCT image reconstruction using arbitrary orbits.
• To address the limitations of conventional iterative reconstruction algorithms in terms of speed and resource requirements.

Main Methods:

• Development of a differentiable shift-variant filtered backprojection neural network optimized for arbitrary trajectories.
• Integration of known operators into the neural network to minimize trainable parameters and enhance interpretability.
• Adaptation of the framework for non-continuous trajectories like circular-plus-arc and sinusoidal paths.

Main Results:

• Achieved over 97% reduction in computation time compared to conventional iterative algorithms.
• Demonstrated superior or comparable image quality with a 38.6% reduction in mean squared error.
• Showcased improvements in image quality metrics, including a 7.7% increase in peak signal-to-noise ratio and a 5.0% increase in structural similarity index measure.
• Validated flexibility and robustness across diverse scan geometries.

Conclusions:

• The novel neural network method enables real-time, high-quality CBCT reconstructions for customized orbits.
• Represents a significant advancement for interventional medical imaging, especially for robotic C-arm CT systems.
• Offers a transformative solution for clinical applications demanding computational efficiency and imaging precision.