The prototype
We built the offline registration core for a spine-navigation concept. The prototype detected two rigid tracking tools in a model-spine CT scan, represented selected vertebral edges, and calculated the transform needed to relate tracker motion to bone geometry.
SYSTEM ARCHITECTURE
Detect the tracking tools
DICOM volume → contour filtering → persistent sphere centers → two tracker framesCircularity, size, and across-slice persistence isolated eight fiducial spheres and grouped them into two labeled four-ball trackers.
Represent the vertebrae
Orthogonal CT projections → user-selected edges → bone line modelsWhen automated segmentation proved unreliable, we delivered a user-guided projection tool that preserved the measurement workflow.
Register the coordinate frames
tracker frames + selected bone lines → nearest tracker-line pairing → rigid transform → relative bone anglePairing each line with its nearest tracker established the offline mapping needed to convert later tracker poses into vertebral measurements.
What we built
- STAGE 01
Detect the fiducials
Processed the DICOM stack and detected fiducial candidates using contour geometry and persistence across neighboring slices.
- STAGE 02
Recover tracker frames
Grouped the eight recovered sphere centers into two rigid tracker frames using coplanarity and known marker geometry.
- STAGE 03
Choose a usable bone model
Tested 3D V-Net and U-Net segmentation, then retained a manual projection-and-line workflow when the learned masks were not reliable.
- STAGE 04
Register tracker and bone geometry
Paired tracker frames with selected vertebral edges and calculated the rigid transform and relative angle between them.
Design decisions
Reliable geometry mattered more than a plausible image
A bright circular contour was not enough to identify a fiducial. Persistence across slices, coplanarity, and known inter-ball geometry turned noisy detections into labeled coordinate frames.
The manual fallback was the honest result
The learned segmentation looked promising but failed its loss metric. We kept the user-guided line tool as the working path; live tracking, repeatability, and clinical accuracy remained future integration and validation work.
What worked
- The 192-slice test volume produced eight sphere centers grouped into two four-ball tracking tools.
- The functional workflow saved two vertebral line models, paired them with the trackers, and produced a 15.55° sample angle.
- The notebooks and transformation plots demonstrate the offline CT-to-tracker registration path from image input to geometric output.
Takeaway
Making every coordinate frame explicit turned a difficult vision problem into smaller, testable interfaces.
A working manual path was more valuable than an automated segmentation that looked plausible but failed its validation metric.