pipeline:

method: MedSAM2 (github)

observation: a bad tooth is part of a good tooth. a tooth is a simple surface when viewd from center-of-mass

method: HEALPix NESTED ordering can translate an stl surface to a 1-d array -> only have to do 1d masked autoencoding

*HEALPix: see attached pdf (Górski 2005)

observation: if a corrupted tooth have a sufficiently large portion intact, our MAE can render a full tooth

subtracting from rentered tooth the corrupted GT gives us a Prosthodontics treatment

method: MFEM (github)

the advantage of this pipeline is its economy on the scarce bad-tooth data, and leverage abundant full-teeth datasets and models.

a pre-experiment using only existing stl- format single tooth annotation yields a working tooth MAE and such data can be obtained via [full mouth CBCT + tooth segmentation], the latter was actively researched

**this codebase is only a pre-experiment validating the 2nd step. **