Cone-beam CT on mobile C-Arm systems enables three-dimensional intraoperative imaging during surgical procedures, improving guidance particularly in orthopedic and trauma surgery. However, metallic implants cause severe image artifacts that obscure critical anatomical structures and compromise surgical decision-making. While post-processing metal artifact reduction techniques exist, they struggle in severe cases and can even introduce secondary artifacts. This motivates trajectory optimization strategies that minimize artifact formation at the acquisition stage. This thesis presents novel methods for metal artifact avoidance in intraoperative cone-beam CT imaging through trajectory optimization. First, geometry-aware neural network architectures are developed to enhance scene understanding from X-ray scout views by embedding known imaging system geometry as differentiable operators within deep learning models. Two such operators are utilized: a differentiable backprojection operator that enables volumetric metal segmentation from few given projection images, and an epipolar view translation operator that leverages geometric constraints between views to improve 2D segmentation performance. These methods provide the basis for subsequent trajectory planning. Second, an interactive framework for intraoperative metal artifact avoidance is presented that computes 3D localized and calibrated artifact predictions from scout images to guide trajectory optimization. A physics-based artifact severity model enables absolute artifact strength estimation. Through a real-time artifact visualization overlayed onto two previously acquired scout views, surgeons can interactively optimize C-arm trajectories before acquisition, prioritizing clinically relevant anatomical regions. A cadaver study demonstrates significant image quality improvements and additional enhancements when factoring in an explicit imaging task from clinical context. This work represents the first metal artifact avoidance system validated in a realistic surgical scenario. Third, trajectory optimization is reformulated using parametric object representations instead of voxel-based methods. Ellipsoidal surrogate models fitted to metallic screws enable optimization based on simplified geometric primitives, thereby avoiding computationally expensive raytracing-based methods. An end-to-end system detects keypoints using a deep neural network, computes the detected objects' three dimensional pose and position through triangulation, and optimizes trajectories using analytic ellipsoid perspective projection. Validated on cadaveric data, this method achieves substantial speedup over traditional approaches while operating entirely on CPU, making it suitable for real-time intraoperative deployment. The integration of geometric prior knowledge into deep learning architectures, combined with interactive workflow design and parametric object modeling, establishes a foundation for intelligent, context-aware imaging systems that adapt dynamically to the surgical scene. These contributions advance intraoperative cone-beam CT image quality in the presence of metallic objects, with promising pathways toward clinical adoption and broader applications in interventional imaging.
Maximilian Rohleder (Thu,) studied this question.