Learning to Avoid Poor Images: Towards Task-aware C-arm Cone-beam CT Trajectories

Jan-Nico Zaech, Cong Gao, Bastian Bier, Russell H. Taylor, Andreas K. Maier, Nassir Navab, Mathias Unberath

Abstract

Intraoperative 3D Cone-beam CT (CBCT) is vital for verifying surgical outcomes, such as pedicle screw placement in spinal fusion. However, metal artifacts often degrade image quality, obscuring critical anatomical interfaces. This paper proposes a task-aware trajectory system that uses a deep learning model to predict optimal imaging angles in real-time. By maximizing a physics-based detectability index, the robotic C-arm autonomously avoids "poor" views, resulting in significantly clearer 3D reconstructions and a 19% improvement in feature detectability over conventional circular scans.

Introduction

In spinal fusion surgery, the precise placement of pedicle screws is paramount to avoid nerve damage. While intraoperative CBCT provides 3D visualization, the presence of metal leads to severe artifacts—dark streaks and bright flares—caused by beam hardening and photon starvation. Traditional software-based metal artifact reduction (MAR) often fails when the raw data is fundamentally corrupted by physics. This research shifts the focus from post-processing to acquisition, asking whether the X-ray source can be moved to a better position to avoid these physics-based corruptions entirely.

Methodology

The authors introduce a "task-aware" trajectory system that dynamically adjusts the C-arm's out-of-plane tilt. The core of the system is the detectability index ($d^2$), a physics-based metric that quantifies how well a feature (e.g., a screw) can be visualized given the system's noise and resolution characteristics.

To enable real-time control without a prior 3D model of the patient, a modified VGG-19 Convolutional Neural Network (CNN) was developed. This network is trained to regress the detectability of potential subsequent views based solely on the current 2D X-ray projection. As the C-arm rotates, the CNN identifies the tilt angle that maximizes detectability, and the robotic arm "servos" to this position on-the-fly.

The proposed pipeline: A 2D X-ray projection is processed by a VGG-19 network to predict a detectability heatmap, which then guides the robotic C-arm trajectory.

Experimental Setup

The system was trained on a synthetic dataset comprising over 290,000 images generated from 5 chest CT volumes with simulated pedicle screws. High-fidelity simulations were performed using DeepDRR, which accurately models X-ray physics and metal artifacts. Real-world validation was conducted using a physical phantom (ballistic gel and iron screws) and a Siemens Arcadis Orbic 3D robotic C-arm to evaluate the model's generalizability to clinical hardware.

Results and Discussion

The CNN demonstrated high predictive accuracy, with an average angular error of only 8.35° on synthetic test data. In physical phantom experiments, the task-aware trajectory achieved a 19% increase in predicted detectability compared to standard circular trajectories.

A key insight is that the optimal next view depends primarily on the current view and anatomy (the Markov property), making the system robust to initial positioning. By avoiding "screw overlap"—where X-rays pass through the long axis of the metal—the system preserves raw signal integrity. Qualitatively, reconstructions from the task-aware path showed clearly defined screw threads and reduced "cupping" artifacts, which were otherwise obscured in standard scans.

Conclusion

This work demonstrates the potential of autonomous, physics-informed robotics to improve medical imaging. While current hardware limitations hinder the immediate adoption of complex non-planar trajectories in high-speed scans, the framework provides a foundation for task-specific imaging. Future research will explore more complex surgical tasks and validation on cadaveric specimens.

Summary updated

Abstract

Metal artifacts in computed tomography (CT) arise from a mismatch between physics of image formation and idealized assumptions during tomographic reconstruction. These artifacts are particularly strong around metal implants, inhibiting widespread adoption of 3D cone-beam CT (CBCT) despite clear opportunity for intra-operative verification of implant positioning, e.g. in spinal fusion surgery. On synthetic and real data, we demonstrate that much of the artifact can be avoided by acquiring better data for reconstruction in a task-aware and patient-specific manner, and describe the first step towards the envisioned task-aware CBCT protocol. The traditional short-scan CBCT trajectory is planar, with little room for scene-specific adjustment. We extend this trajectory by autonomously adjusting out-of-plane angulation. This enables C-arm source trajectories that are scene-specific in that they avoid acquiring "poor images", characterized by beam hardening, photon starvation, and noise. The recommendation of ideal out-of-plane angulation is performed on-the-fly using a deep convolutional neural network that regresses a detectability-rank derived from imaging physics.
BibTeX
@inproceedings{zaech2019learning,
  author = {Jan-Nico Zaech and Cong Gao and Bastian Bier and Russell H. Taylor and Andreas K. Maier and Nassir Navab and Mathias Unberath},
  title = {{Learning to Avoid Poor Images: Towards Task-aware C-arm Cone-beam CT Trajectories}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2019},
  url = {https://openreview.net/forum?id=W2j8T8BgANJ},
}
Figure