Hierarchical Classification via Cascading Feature Elimination: Application to Human Phenotype Ontology-Aligned Facial Phenotyping (FaceMesh2HPO)

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Health & Wellbeing — Clinical Care & Medical Practice, Medical Devices & Health Technology, Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

FaceMesh2HPO is a novel framework for classifying facial phenotypic descriptors aligned with the Human Phenotype Ontology (HPO) to support genetic disorder diagnosis. It addresses limitations of existing syndrome-level prediction methods by providing structured descriptions of facial morphology. The framework utilizes 3D facial meshes with 478 automatically detected points from 2D images, combined with age, sex, and ethnicity metadata. A hierarchical, cascading classification pipeline of PointNet-based models is trained along the HPO tree, employing iterative point elimination based on term-specific importance. The best configuration achieved a mean AUROC of 0.750± 0.042 across 107 HPO models in cross-validation, with 3D meshes, facial outline, and metadata consistently improving performance. Parent and "compression" nodes generally outperformed leaf nodes, with top models reaching AUROCs of ≈ 0.89. The system partially transfers to unseen disorders, particularly at parent HPO term levels.

Key takeaway

For clinical geneticists and diagnostic teams evaluating patients with suspected genetic disorders, FaceMesh2HPO offers an interpretable, HPO-aligned tool for facial phenotyping. You can use its web application to upload patient images and obtain structured phenotypic descriptions, which can directly inform differential diagnoses and complement existing syndrome classification systems. This approach provides granular, ontology-linked support, especially for common traits and at coarser ontology levels, enhancing trust and utility in diagnostic workflows.

Key insights

Hierarchical classification of HPO-aligned facial phenotypes using 3D meshes and cascading feature elimination improves diagnostic support and interpretability.

Principles

• 3D facial meshes and metadata enhance phenotype classification accuracy.
• Hierarchical models benefit from larger sample sizes at higher ontology levels.
• Feature elimination reduces complexity and focuses models on salient regions.

Method

Train a hierarchical, cascading PointNet-based classification pipeline on 3D facial meshes, dynamically adjusting architecture and iteratively pruning mesh points via Integrated Gradients based on term-specific importance.

In practice

• Use 3D face meshes, facial outline, age, sex, and ethnicity metadata.
• Apply a point-importance threshold of 0.01 for feature elimination.
• Employ soft labels of 0.05 for negative samples to mitigate uncertainty.

Topics

• Human Phenotype Ontology
• Facial Phenotyping
• Deep Learning
• PointNet
• 3D Face Meshes
• Genetic Disorders Diagnosis
• Feature Elimination

Code references

• hcmlab/FaceMesh2HPO

Best for: Computer Vision Engineer, AI Scientist, Research Scientist, Domain Expert

Related on AIssential

• Hierarchical Classification via Cascading Feature Elimination: Application to Human Phenotype Ontology-Aligned Facial Phenotyping (FaceMesh2HPO) — Hierarchical classification of 3D facial geometry enables interpretable, ontology-linked phenotype classification for clinical diagnosis.
• Single-Stage Hierarchical Rectification for Weakly Supervised Histopathology Segmentation — SSHR proactively purifies features in a single stage, outperforming multi-stage WSSS and reducing training time significantly.
• TraMP-LLaMA: Generative Interpretability with Decoupled Instruction Tuning for Facial Expression Quality Assessment — TraMP-LLaMA provides generative interpretability for facial expression quality assessment by jointly predicting scores and generating textual reports.
• TIGER: Taming Identity, Geometry, and Generative Priors for High-Quality Face Video Restoration — TIGER fuses identity, geometry, and generative priors to restore high-fidelity, identity-consistent face videos efficiently.
• Geometry-Consistent Endoscopic Representations for Image-Guided Navigation via Structured Foundation Model Adaptation — Hierarchy-Aware Geometry-Semantic Adaptation improves monocular endoscopy navigation by integrating geometric supervision and structured foundation model adaptation for robust representations.
• Embracing Intra-Class Heterogeneity for Semi-Supervised Medical Image Segmentation: From Diversity to Precision — Explicitly modeling intra-class heterogeneity with multiple intensity-aligned prototypes improves semi-supervised medical image segmentation precision, especially with limited data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.