IIT Madras · HTIC — 2026
Retiring a thin-plate-spline pipeline from 1989 and proving the replacement, 89 tests deep — including one that checks the output is byte-identical every run.
Shipped — in use as a drop-in replacement
Image slot
BCPD++ — reconstructed bone geometry against donor ground truth
1600×1000 · side-by-side of TPS vs BCPD reconstruction, or the volume comparison chart
Role
Research engineer — reformulation, implementation, validation, release
Year
2026
Stack
- C++
- Bayesian CPD
- Python
- NumPy
- CMake
The problem
Bone synthesis at HTIC ran on thin-plate splines, a formulation published in 1989 and never revisited. TPS interpolates landmarks, but it has no notion of the body it is deforming — it treats the bone as a rubber sheet pinned at a handful of points. In practice it hit its landmarks only approximately and shrank the reconstructed bone by as much as 65% of donor volume. A geometry that wrong cannot carry a surgical plan, no matter how neatly it fits the landmarks it was given.
Approach
Reformulated the whole stage as Bayesian Coherent Point Drift: probabilistic point-set registration with a motion-coherence prior, so the bone deforms like a body rather than like a sheet.
Split it into two stages — nine anatomical landmarks drive a whole-bone estimate, then the probed surface refines that estimate where the surgeon actually has data.
Released it as drop-in replacement executables behind the existing interface, so adoption required no changes to any calling code.
Wrote 89 tests, including a byte-identical determinism test: same input, same bytes out, every run, every machine. In a pipeline that ends at a cut, reproducibility is a safety property rather than a convenience.
Results
- landmark residual
1.8×10⁻¹³ mm
landmark residual
0.0467 mm under TPS
- volume deviation from donor
2.7%
volume deviation from donor
65% shrinkage under TPS
- better surface morphing
43%
better surface morphing
- RMSE vs ~20,000 swept points
0.24 mm
RMSE vs ~20,000 swept points
0.5 mm accuracy gate
Validated against roughly 20,000 surface points swept by technicians on donor specimens, measured against the team's 0.5 mm accuracy gate.
Image slot
TPS versus BCPD reconstruction compared against the donor bone
1600×1000
Image slot
Test suite output showing the determinism check passing
1600×1000
What it came to
The landmark residual moving from 0.0467 mm to 1.8×10⁻¹³ mm is really a statement that the new formulation solves the constraint exactly instead of approximating it. The volume figure is the one that changed clinical confidence: a bone that keeps its donor volume to within 2.7% is a bone you can plan against.