AI registers knee bones from single X-ray with 758-patient model
No initialization needed: PnP+RANSAC solves 6-DoF pose from one radiograph
Recovering the 6-DoF pose of knee bones from a single X-ray given a pre-operative CT scan is a long-standing challenge. Classic methods align rendered bone silhouettes to image edges, which suffer from depth ambiguity, while recent differentiable-rendering approaches refine pose iteratively but have a narrow capture range and high per-iteration cost. Both require per-patient optimization and are fragile under a single view.
Rembert Daems and colleagues propose learning an amortized, subject-agnostic dense 2D-3D correspondence using a shared-weight model per bone, trained across 758 patients. The model is supervised solely by projection geometry and outputs transmission-aware correspondence (not tied to a single surface). Pose is then solved in closed form via global, initialization-free PnP+RANSAC. The representation is anatomically semantic: a simple classifier can read a landmark's region from its embedding, and features separate kneecap, femur, and tibia without bone labels. On a large single-institution cohort, the model generalizes well to held-out patients, enabling low-dose quantitative joint assessment without repeat CT or fixed biplanar rigs.
- Model trained on 758 patient CT-X-ray pairs generalizes to unseen patients without per-subject optimization
- Transmission-aware dense correspondence handles X-ray transparency, unlike surface-based silhouette methods
- Closed-form pose estimation via PnP+RANSAC eliminates need for initialization or costly differentiable rendering
Why It Matters
Enables quantitative knee joint geometry from routine low-dose X-rays, reducing radiation and need for repeat CT scans.