Introduction
In spinal fusion surgery, the precise placement of pedicle screws is critical to avoid nerve damage. Intraoperative C-arm Cone-beam CT (CBCT) is the standard for verification; however, the presence of metal screws often leads to severe image artifacts. These artifacts, caused by phenomena like beam hardening and photon starvation, can obscure the very anatomy surgeons need to see. This paper presents a novel deep learning approach to mitigate these artifacts by adjusting the C-arm trajectory in real-time.
The Challenge of Metal Artifacts
Standard CBCT scanners follow a fixed circular path. When the X-ray beam aligns with the long axis of a dense metal screw, the resulting data is highly corrupted, leading to "blooming" and dark streaks in the 3D reconstruction. While post-processing metal artifact reduction (MAR) algorithms exist, they often introduce secondary errors. This research proposes a proactive solution: optimizing the scan trajectory on-the-fly to avoid problematic viewing angles.
Methods: Online Trajectory Adjustment
The authors developed an online pipeline that adjusts the scanner's "out-of-plane" tilt based on live X-ray projections.
The core of this system is a VGG-style convolutional neural network (CNN) acting as a quality regressor. For each 2D projection, the CNN predicts a "detectability index" ($d^2$)—a metric for task-specific image quality—for 11 potential next positions. Unlike previous methods that require a pre-operative 3D model, this approach is model-free at runtime, as the network implicitly learns the relationship between 2D features and 3D reconstruction quality. To ensure mechanical feasibility, a regularization term is included to prevent erratic robotic movements.
Experimental Evaluation
The model was trained primarily on synthetic data generated using the DeepDRR simulator and chest CT scans from the Cancer Imaging Archive (TCIA). Realistic titanium screws and noise were added to the simulations. Validation was performed using both synthetic tests and a semi-anthropomorphic phantom containing metal screws.
Key Results
The optimized trajectories significantly outperformed standard circular scans:
- Artifact Reduction: In synthetic tests, the "blooming" effect (measured by Full Width at Half Maximum) was reduced from ~6.38 mm in circular scans to ~3.65 mm, closely matching the ground truth of 3.92 mm.
- Improved Detail: In phantom experiments, fine screw threads that were blurred in circular scans became clearly visible in the optimized reconstructions.
- Task-Awareness: The system naturally learned to avoid the "long axis" of screws, prioritizing angles that maximize the detectability of critical features.
Limitations and Future Directions
Despite the promising results, some challenges remain. The current model occasionally exhibits "jitter" or oscillatory behavior because it lacks temporal context (the Markov property). Future iterations may incorporate Recurrent Neural Networks (RNNs) to provide memory and smoother transitions. Additionally, the transition from simulation to real clinical environments requires further domain adaptation to account for complex human anatomy and the mechanical constraints of heavy robotic C-arms.
Summary updated