Depth Anything 3 + ZEISS microscopes reconstruct 3D surgical scenes with ~1mm accuracy
Researchers used a pretrained AI and standard OR microscope footage to get metric 3D surfaces of brain surgery sites.
Researchers at Inselspital Bern and the University of Bern have demonstrated that metric 3D surface reconstruction of neurosurgical operative corridors is possible using standard monocular operating microscope images combined with microscope pose data from a navigation system. In their phantom-based study, they used a ZEISS Pentero 800 microscope integrated with Brainlab Cranial Navigation, capturing standard composite video output alongside synchronous microscope poses. Depth estimation was performed with the pretrained Depth Anything 3 model—no fine-tuning was required—and point clouds were fused into meshes via Poisson surface reconstruction.
Two aneurysm training phantoms were tested: Phantom A (deep surgical corridor) yielded an accuracy of 1.95 ± 1.70 mm to 2.33 ± 2.15 mm, while Phantom B (directly exposed surface) achieved 1.02 ± 0.93 mm to 1.52 ± 1.21 mm. Larger image sets mainly improved completeness without significantly changing accuracy. The results show technical feasibility in a controlled setting, paving the way for automated quantification of operative exposure, image fusion with preoperative scans, and characterization of working spaces for surgical instrumentation.
- Depth Anything 3 foundation model used without any task-specific fine-tuning for depth estimation from monocular microscope images.
- Reconstruction accuracy of 1–2 mm achieved on phantom surgical scenes using ZEISS Pentero 800 microscope and Brainlab navigation data.
- Method works with standard composite video output, requiring only pre-operative intrinsic/extrinsic calibration plus synchronized microscope pose.
Why It Matters
Enables objective, automated 3D mapping of surgical exposure from existing OR equipment, potentially improving safety and planning in neurosurgery.