Objective: Accurate placement of external ventricular drains (EVDs) is achieved in only approximately 67–74% of cases using the conventional freehand technique. Augmented reality (AR) offers the potential to improve this by providing real-time, patient-specific anatomical guidance. This thesis evaluates whether CT-based anatomical landmark registration using the Lumi AR workflow is sufficiently accurate, robust, and feasible to support and eventually improve EVD placement. This also includes exploring the clinical acceptability of AI-generated landmarks to streamline the workflow.
Methods: Two studies were performed. First, four clinicians assessed the accuracy of AI-generated anatomical landmarks on CT-derived 3D models, with adjustment rates and interobserver agreement quantified. Second, a prospective pilot study in the operating room (OR) was conducted using the Lumi AR workflow on the HoloLens 2 to perform point-based registration with manually annotated landmarks. The primary outcome was target registration error (TRE); secondary outcomes included fiducial registration error (FRE), visual accuracy ratings, registration time, system robustness and workflow feasibility.
Results: AI-generated landmarks required adjustment in 22.9% of cases (95% CI, 19.1–27.1%), with high median partial interobserver agreement (100.0%, IQR 25.0%) but only moderate mean unanimous agreement (61.0%, 95% CI 51.4–69.7%; Fleiss’ kappa = 0.42). In the OR pilot (n=11), the mean TRE at the nasion was 4.9 mm (SD, 2.1 mm). For fiducial validation points, mean TREs were 7.4 mm (SD, 1.7 mm) and 4.9 mm (SD, 1.9 mm). The mean FRE was non-inferior to that reported in a previous phantom study, visual accuracy ratings indicated good perceived alignment, and registration was completed in five minutes on average. Workflow interruptions were primarily due to hardware instability, including three critical failures.
Discussion & Conclusion: AI-generated anatomical landmarks are not yet sufficiently reliable for clinical use in high-stakes scenarios such as EVD placement. In contrast, point-based registration with manually annotated landmarks, using the Lumi AR workflow, proved clinically feasible and achieved an accuracy that is likely acceptable for EVD guidance. However, system robustness remains a key limitation, with AR hardware instability representing the primary obstacle to clinical implementation. Additional limitations include the small pilot sample size, which restricts generalisability, and the variability of soft-tissue surface landmarks. While further advances in AR hardware and validation in larger cohorts are required, these findings indicate that CT-based anatomical landmark registration using AR shows clear potential for guiding future EVD placements.