Abstract
PURPOSE: Endoscopic transsphenoidal surgery for pituitary tumors is hindered by limited visibility and maneuverability due to the narrow nasal corridor, increasing the risk of complications. To address these challenges, we present a pipeline for 3D reconstruction of the sellar anatomy from monocular endoscopic videos to enhance intraoperative visualization and navigation. METHODS: Data were collected through a user study with trainee surgeons, and the procedure was conducted on 3D printed, anatomically correct phantom devices. To overcome limitations posed by the uniform, textureless surfaces of these devices, learned feature detectors and matchers were leveraged to extract meaningful information from the images. The matched features were reconstructed using COLMAP, and the resulting surfaces were evaluated using the iterative closest point algorithm against the CAD ground-truth surface of the printed phantoms. RESULTS: Most methods resulted in accurate reconstructions with moderate variability in cases with high blur or occlusions. Average RMSE values of 0.33 mm and 0.41 mm, for the two best methods, Dense Kernelized Feature Matching and SuperPoint with LightGlue, respectively, were obtained in the surface registrations across all test sequences, with a significantly higher computation time for Dense Kernelized Feature Matching. CONCLUSION: The proposed pipeline was able to accurately reconstruct anatomically correct 3D models of the phantom devices, showing potential for the use of learned feature detectors and matchers in real time for AR-guided navigation in pituitary surgery.