Abstract
The recent curation of large-scale databases with 3D surface scans of shapes has motivated the development of tools that better detect global patterns in morphological variation. Studies, which focus on identifying differences between shapes, have been limited to simple pairwise comparisons and rely on prespecified landmarks (that are often known). We present SINATRA, the first statistical pipeline for analyzing collections of shapes without requiring any correspondences. Our novel algorithm takes in two classes of shapes and highlights the physical features that best describe the variation between them. We use a rigorous simulation framework to assess our approach. Lastly, as a case study we use SINATRA to analyze mandibular molars from four different suborders of primates and demonstrate its ability recover known morphometric variation across phylogenies.
Introduction
In fields like evolutionary biology and medical imaging, researchers often need to identify the specific physical features that distinguish one group of shapes from another. Traditionally, this has required geometric morphometrics, a process dependent on "landmarks"—specific anatomical points that must be manually identified on every sample.
This manual approach is limited by human bias, significant information loss (as complex 3D scans are reduced to a few dozen points), and topological constraints that make it difficult to compare shapes with different features, such as a bone with a unique protrusion. To overcome these hurdles, we introduce SINATRA, a landmark-free statistical pipeline that automatically identifies and ranks the physical regions most responsible for distinguishing between classes of 3D meshes.
The SINATRA Pipeline
SINATRA processes 3D data through four distinct stages to transform raw geometry into interpretable statistical evidence:
- Topological Summarization: Instead of landmarks, the pipeline uses the Euler Characteristic (EC) transform. By sweeping a plane through a 3D shape in multiple directions and recording the Euler characteristic (Vertices - Edges + Faces) at each step, the algorithm converts complex 3D objects into 1D curves that capture their essential geometry.
- Statistical Classification: These curves are fed into a Gaussian Process Classification (GPC) model, which learns to distinguish between groups (e.g., Species A vs. Species B).
- Feature Selection: The model employs an information-theoretic metric based on Kullback-Leibler Divergence (KLD) to determine which parts of the topological curves are most statistically significant for the classification.
- Reconstruction: Because the EC transform is invertible, SINATRA maps these high-association features back onto the original 3D mesh, creating a heatmap of the physical regions that differentiate the groups.
Key Results and Biological Validation
The authors validated SINATRA using both simulated data (perturbed spheres and caricatured teeth) and a real-world dataset of 59 primate mandibular molars.
- Superior Sensitivity: In simulations, SINATRA significantly outperformed baseline methods as shape complexity increased, accurately identifying class-specific indentations that other tools missed.
- Evolutionary Insights: In the primate study, the pipeline successfully identified the paraconid—a specific high cusp on the teeth of Tarsius primates—as the primary feature distinguishing them from other suborders. This demonstrates the tool's ability to recover known evolutionary traits without prior human guidance.
- Interpretability: By projecting statistical weights back into 3D space, SINATRA turns abstract machine learning results into a visual tool that biologists can use to generate new hypotheses.
Conclusion
SINATRA represents a major step forward in 3D shape analysis by providing a landmark-free, statistically rigorous, and interpretable framework. While currently focused on binary classification, future iterations aim to extend the pipeline to multi-class problems and continuous trait correlations. By allowing the data to "speak for itself," SINATRA enables the discovery of morphological patterns that might be overlooked by traditional human-centric methods.
Reference: Wang, B., et al. (2021). A statistical pipeline for identifying physical features that differentiate classes of 3D shapes. The Annals of Applied Statistics, 15(2), 638–661.
Summary updated