FaceMesh2HPO is a framework designed to support clinical diagnosis by classifying facial phenotypic descriptors aligned with the Human Phenotype Ontology (HPO). Utilizing annotations from 124 clinicians across 10 disorders (107 HPO terms) along with non-syndromic controls, 3D facial meshes with 478 landmarks were generated from 2D images. A hierarchical PointNet-based pipeline was employed for cascading classification and feature elimination.
The best models, which integrated 3D meshes, facial outlines, and demographic metadata, achieved AUROC scores ranging from ~0.55 to ~0.89, with better performance observed at parent nodes compared to leaf terms. External validation indicated variable generalizability across disorders. Results demonstrate that hierarchical modeling of 3D facial geometry enables interpretable, ontology-linked phenotype classification, though performance on rare leaf terms remains limited. Enhanced data diversity and feature selection strategies are necessary to improve robustness and clinical utility.
Blogger's Review: FaceMesh2HPO showcases the potential of 3D facial geometry in clinical diagnosis, particularly in phenotype classification. While the model performs well at parent nodes, its limitations with rare phenotypes highlight that data quality and diversity are crucial for improving classification performance. Future research should focus on effectively integrating diverse features to enhance the model's generalizability and clinical relevance.