[Paper Review] A Learning-based Method for Online Adjustment of C-arm Cone-Beam CT Source Trajectories for Artifact Avoidance

Mareike Thies, Jan-Nico Zäch+5 more|arXiv (Cornell University)|Jan 1, 2020

Spinal Fractures and Fixation Techniques33 references30 citations

TL;DR

This paper proposes a deep learning method for real-time, online adjustment of C-arm cone-beam CT (CBCT) source trajectories during spinal surgery to reduce metal artifacts from pedicle screws. Using a convolutional neural network that regresses a view-specific quality index from live fluoroscopic projections, the system dynamically selects optimal next views to avoid data inconsistencies, resulting in significantly improved reconstruction quality with reduced artifacts compared to standard circular trajectories.

ABSTRACT

During spinal fusion surgery, screws are placed close to critical nerves suggesting the need for highly accurate screw placement. Verifying screw placement on high-quality tomographic imaging is essential. C-arm Cone-beam CT (CBCT) provides intraoperative 3D tomographic imaging which would allow for immediate verification and, if needed, revision. However, the reconstruction quality attainable with commercial CBCT devices is insufficient, predominantly due to severe metal artifacts in the presence of pedicle screws. These artifacts arise from a mismatch between the true physics of image formation and an idealized model thereof assumed during reconstruction. Prospectively acquiring views onto anatomy that are least affected by this mismatch can, therefore, improve reconstruction quality. We propose to adjust the C-arm CBCT source trajectory during the scan to optimize reconstruction quality with respect to a certain task, i.e. verification of screw placement. Adjustments are performed on-the-fly using a convolutional neural network that regresses a quality index for possible next views given the current x-ray image. Adjusting the CBCT trajectory to acquire the recommended views results in non-circular source orbits that avoid poor images, and thus, data inconsistencies. We demonstrate that convolutional neural networks trained on realistically simulated data are capable of predicting quality metrics that enable scene-specific adjustments of the CBCT source trajectory. Using both realistically simulated data and real CBCT acquisitions of a semi-anthropomorphic phantom, we show that tomographic reconstructions of the resulting scene-specific CBCT acquisitions exhibit improved image quality particularly in terms of metal artifacts. Since the optimization objective is implicitly encoded in a neural network, the proposed approach overcomes the need for 3D information at run-time.

Motivation & Objective

• To address the critical clinical challenge of metal artifacts in intraoperative CBCT during spinal fusion surgery, which compromise screw placement verification.
• To improve reconstruction quality by proactively selecting x-ray views that are most consistent with the idealized image formation model, minimizing data inconsistencies.
• To eliminate the need for 3D volumetric information at scan time by learning task-specific trajectory adjustments from 2D projection images alone.
• To enable real-time, patient-specific trajectory optimization using a deep neural network trained on realistic simulated data.
• To demonstrate feasibility and improvement over standard circular CBCT scans using both simulated and real phantom data.

Proposed method

• A convolutional neural network (VGG-19) is trained to predict a view-dependent quality index from a single live fluoroscopic projection image.
• The network outputs a recommendation for the next optimal C-arm source angle (in-plane and out-of-plane) based on the current projection, enabling online, incremental trajectory updates.
• The method uses a reinforcement learning-inspired policy to select views that minimize inconsistencies with the idealized reconstruction model, particularly avoiding beam hardening and photon starvation effects.
• Training data is generated from realistic Monte Carlo simulations of x-ray projections with titanium pedicle screws, capturing realistic scatter, beam hardening, and noise.
• The network is evaluated retrospectively on both simulated and real CBCT data from a semi-anthropomorphic phantom to assess reconstruction quality.
• Trajectories are optimized to avoid views along screw long axes and overlapping screw projections, which are known to cause severe artifacts.

Experimental results

Research questions

• RQ1 Can a deep learning model accurately predict the quality of future x-ray views based solely on a single current projection image during a CBCT scan?
• RQ2 Can online, real-time adjustment of the C-arm trajectory using neural network predictions reduce metal artifacts in CBCT reconstructions compared to standard circular trajectories?
• RQ3 To what extent can a model trained on simulated data generalize to real-world CBCT acquisitions from a phantom?
• RQ4 How do the resulting non-circular trajectories compare to conventional circular orbits in terms of artifact reduction and anatomical detail preservation?
• RQ5 Can the method improve detectability of screw cortical breaches without requiring 3D information at scan time?

Key findings

• The proposed method successfully reduces metal artifacts in CBCT reconstructions by dynamically adjusting the C-arm trajectory to avoid views that cause severe data inconsistencies.
• The convolutional neural network achieves robust and accurate predictions of view quality across varying noise levels and initialization angles, even when generalizing from simulation to real data.
• Realistic simulated data training enables the network to generalize to real phantom acquisitions, with reconstructed screw shapes and threads showing marked improvement over standard circular scans.
• Trajectories predicted by the network avoid high-artifact views—particularly those along screw long axes and with overlapping screws—leading to more consistent and accurate reconstructions.
• The method demonstrates that task-aware, scene-specific trajectory planning using only 2D projection data can significantly improve image quality without requiring volumetric data at scan time.
• Retrospective evaluation confirms that even a slight tilt from the standard circular plane (parallel to screw axes) leads to substantial artifact reduction, suggesting immediate clinical feasibility.