Metal Artifact Avoidance
Cone-beam CT on mobile C-arm systems lets surgeons see in 3D during procedures, which is especially valuable in orthopedic and trauma surgery. But metallic implants cause severe image artifacts that obscure exactly the anatomical structures surgeons need to see, undermining surgical decisions. Post-processing techniques to reduce these artifacts exist, but they struggle with severe cases and can even introduce new artifacts of their own. This thesis instead tackles the problem at the source: optimizing how the scan is acquired in the first place to avoid creating the artifact.
The first contribution builds geometry-aware neural network architectures that improve scene understanding from X-ray scout views, embedding the known imaging system geometry directly into the deep learning models as differentiable operators. Two such operators are used: a differentiable backprojection operator that segments metal in 3D from just a few projection images, and an epipolar view translation operator that uses the geometric relationship between two views to improve 2D segmentation. Together, these form the foundation for the trajectory planning that follows.
The second contribution is an interactive framework that computes localized, calibrated 3D artifact predictions from scout images to guide trajectory planning in real time, backed by a physics-based model that estimates absolute artifact strength. Surgeons see a live overlay of predicted artifacts on two scout views and can interactively adjust the C-arm’s planned path before acquisition, prioritizing the anatomical regions that matter clinically. A cadaver study confirms meaningful image quality improvements, with further gains when the system accounts for the specific clinical imaging task at hand, making this the first metal-artifact-avoidance system validated in a realistic surgical setting.
The third contribution reformulates trajectory optimization around parametric object representations rather than voxel-based methods. Metallic screws are approximated with fitted ellipsoidal surrogate models, replacing expensive raytracing with simplified geometric primitives. An end-to-end system detects keypoints with a neural network, triangulates their 3D pose and position, and optimizes the trajectory using analytic ellipsoid perspective projection. Validated on cadaver data, this approach runs substantially faster than traditional methods, entirely on CPU, making it practical for real-time intraoperative use.
Together, embedding geometric prior knowledge into deep learning, an interactive workflow, and parametric object modeling lay the groundwork for imaging systems that adapt intelligently to the surgical scene in real time, advancing intraoperative CBCT image quality around metal implants with a clear path toward clinical adoption.