A new reference-invariant consensus template generation method in ALPACA

Automated landmarking transfers anatomical landmarks from a reference specimen onto many targets, greatly increasing analytical throughput. However, this procedure needs to be bootstrap using an initial sample. An arbitrary or atypical choice imprints a reference-of-origin bias that propagates through the pseudo-landmarks, the resulting morphospace, and the downstream template selection, a risk that is difficult to avoid for large datasets whose variation is not yet understood.

We replace the fixed reference with an iterative consensus atlas, warped over a few iterations toward the Procrustes mean shape of all similarity-aligned specimens. We evaluated it on a 62-strain Mus musculus skull panel by running both the original fixed-reference pipeline and the new consensus pipeline 62 times each, using every specimen in turn as the bootstrap. We compared atlas convergence, inter-atlas similarity, morphospace reproducibility, reference-choice variance of pairwise Procrustes distances, downstream k-means selection stability, and leave-one-out out-of-sample fit, and tested generalisation on great-ape datasets of differing sampling balance.

The consensus atlas converged within a few iterations and was far less sensitive to the starting specimen than the fixed reference. It produced more reproducible morphospaces (mean RV 0.960 versus 0.944), reduced the reference-of-origin variance of pairwise distances by a median of about 60%, drew downstream template selections from a smaller and more consistent pool of specimens, and fit held-out specimens more closely in all 62 strains. On the great-ape data the atlases agreed closely in surface geometry, but the downstream morphospace became reference-dependent when the sample was taxonomically imbalanced, and a smaller balanced subset outperformed the larger imbalanced one.

Iterative consensus atlas building removes a persistent bias from automated landmarking and yields reference-invariant, reproducible results, with sampling balance mattering more than absolute sample size. Because the atlas stabilises quickly, it can be built from a small balanced subset while the remaining specimens are simply landmarked against it, a practical route to scaling reference-invariant landmarking. The method is implemented in ALPACA within SlicerMorph, with a mock library enabling headless use on HPC.