New AI infers hidden skull landmarks from face scans with 3mm accuracy
Trained on 240 CT scans, this model reads 21 skeletal landmarks from soft tissue alone...
A new study tackles a key question in 3D facial analysis: can CT-defined internal skeletal landmarks be inferred from external soft-tissue geometry? Using 240 clinical CT scans from two hospitals, the researchers built a locked retrospective protocol pairing CT-derived external soft-tissue point clouds with 21 skeletal landmarks and three visible soft-tissue landmarks. Their integrated hierarchical point-cloud model achieved a mean radial error of 2.97 mm on skeletal landmarks and 3.03 mm on deep or surface-invisible landmarks in 40 held-out patients. Patient-mismatch controls supported patient-specific signal beyond a fixed population configuration or global similarity, while coverage ablations showed dependence on non-anterior geometry. Optical-transfer diagnostics revealed substantial coverage-related and global-configuration components, though deployable optical inference remains unresolved. The results affirm the controlled feasibility of hidden skeletal landmark inference and lay groundwork for future work. The paper was accepted to the AI4M3D Workshop at ECCV 2026.
- Achieves 2.97mm mean radial error on 21 skeletal landmarks, 3.03mm on deep/invisible points
- Trained on 240 clinical CT scans from two hospitals with locked retrospective protocol
- Accepted at AI4M3D Workshop at ECCV 2026; optical inference still an open challenge
Why It Matters
Could eliminate routine CT radiation for orthodontic and surgical planning by predicting bone structure from surface scans.