Immersive Virtual Reality for 3D Cephalometric Landmarking Across Low-Dose and Ultra-Low-Dose CBCT: A Pilot Comparison with a Conventional Computer Interface
Jorma Järnstedt, Helena Mehtonen, Jari Kangas, Hanna Naukkarinen, Kimmo Ronkainen, John Mäkelä, Sakarat Nalampang, Phattaranant Mahasantipiya, Arnon Charuakkra, Wannakamon Panyarak, Irina Rinta-Kiikka, Roope RaisamoBackground: Three-dimensional cephalometric landmarking provides the spatial reference framework for computer-aided surgical simulation in craniomaxillofacial (CMF) surgery. Conventional workflows rely on two-dimensional computer interfaces (CIs), yet imaging data are inherently volumetric, and repeated CBCT imaging creates pressure to minimise patient radiation exposure. Immersive virtual reality (VR) offers a more intuitive environment for spatial tasks, yet the feasibility of ultra-low-dose (uLD) protocols for cephalometric landmarking in CI and VR remains unevaluated. This pilot study evaluated accuracy, reproducibility and workload across low-dose (LD) and uLD CBCT protocols in both environments. Methods: Four CMF radiologists placed ten 3D cephalometric landmarks on 20 CBCT datasets across three rounds. Round 1 used deep learning-predicted coordinates as ground truth. Rounds 2–3 assessed blind reproducibility. Workload was evaluated using the NASA Task Load Index. Results: Median Round 1 accuracy was 1.00 mm in both environments; mean CI accuracy was 1.05 mm and mean VR accuracy was 2.03 mm after outlier exclusion. Log analysis identified controller slips and software logging errors as a primary VR outlier source (22/793, 2.77%). Reproducibility was higher in CIs; VR medians remained clinically relevant. Dose level had no meaningful effect. The NASA-TLX showed decreasing VR workload across sessions, with AI-guided landmarking associated with the lowest mental demand. Conclusions: VR-based landmarking is feasible: its median accuracy was comparable to that of CIs and radiologists responded positively. Reproducibility was lower in VR, attributable to software constraints rather than the visualisation modality. The uLD protocol performed comparably to LD, supporting dose optimisation. Integration of AI assistance within VR represents the most promising direction for further development.