Automated 3D Landmarking Pipeline
A private, validated scientific application that connects bone segmentation, conditioned surface registration, multi-template landmark transfer and geometric-morphometric analysis in one resumable workflow. It automates the repetitive work while keeping coordinate integrity, uncertainty and review decisions visible.
- lower mean error with multi-template consensus
- 49%
- faster deformable registration
- 61×
- leave-one-out validation registrations
- 156
01 · Problem
What needed solving
Placing anatomical landmarks consistently across hundreds of 3D specimens takes weeks of skilled work and introduces observer variation. Existing tools cover isolated stages, leaving researchers to move files, coordinate systems and assumptions between segmentation, registration and analysis software.
Years of landmarking, 3D scanning and supporting morphometric research made the bottleneck clear: the valuable expert decision should be review and interpretation, not repeating the same manual transfer hundreds of times.
02 · Constraints
The difficult parts
- Register clean templates against segmentation-derived meshes that contain interior geometry and acquisition artefacts.
- Preserve coordinate frames and physical units across CT volumes, meshes and landmark files instead of assuming they agree.
- Make deformable registration fast enough for unattended batches without changing numerical behaviour across CPU and GPU.
- Turn model disagreement into an honest review queue rather than presenting every automated placement as equally reliable.
- Keep large batches resumable so a corrupt or incompatible specimen cannot invalidate completed work.
03 · Approach
How the work was structured
- Connected automatic bone segmentation, surface conditioning and morphometric analysis behind one desktop workflow.
- Combined global feature matching, rigid refinement and deformable point-set registration to transfer landmarks from several annotated templates.
- Reimplemented Coherent Point Drift in PyTorch and tested CPU–GPU equivalence before using acceleration in production batches.
- Used a geometric-median consensus across templates and retained their disagreement as a per-landmark signal for triage.
- Added explicit checks for coordinate frames, declared units and whether landmarks fall on bone in the corresponding CT volume.
- Wrote each completed result immediately and made reruns skip finished specimens so long jobs are safe to resume.
04 · Validation
What can be claimed now
Leave-one-out evaluation used an anonymised reference set of 13 fully annotated specimens, producing 156 independent registrations against withheld hand-placed ground truth. Multi-template consensus reduced mean landmark error by 49% relative to the single-template rigid baseline. The PyTorch deformable stage matched the reference implementation to approximately one millionth of the object span.
GPU acceleration reduced the deformable step from 118.7 seconds to 2.0 seconds at the default working size—a 61× speedup—and changed an estimated 300-specimen registration batch from roughly 46 hours to about 1.3 hours. The complete system now produces analysis-ready coordinates, uncertainty-ranked review and direct hand-off to the surrounding GMM application.
05 · Learning
What the project clarified
- More automation is useful only when failure signals and the path to human review remain visible.
- Real segmented surfaces exposed conditioning failures that clean synthetic tests did not reproduce.
- Template coverage across the expected shape range matters more than simply adding more templates.
- Unit errors can survive shape normalisation and create convincing but entirely artefactual allometric results.
- Negative results belong in the design record: a tested statistical shape prior did not improve consensus and was not promoted as a success.