From Screening to Differential Diagnosis: Benchmarking 3D Facial Representations Across Craniofacial Conditions

Rare genetic disorders often present with highly heterogeneous physical features: some have distinctive craniofacial characteristics, whereas others produce subtler or more variable facial differences, making phenotype-based diagnosis challenging. Three-dimensional (3D) facial scanning is a noninvasive approach that captures facial shape and surface geometry, providing quantitative information on craniofacial morphology beyond conventional 2D photographs. Here, we investigated how well 3D facial representations support clinical tasks ranging from broad screening to fine-grained differential diagnosis without landmark registration, dense correspondence, or template fitting. We benchmarked four 3D models, including PointNet, PointNet++, DGCNN, and a geometric baseline, using 26,990 facial scans from 18,785 subjects across four prediction tasks. Across models, performance declined as greater diagnostic specificity was required. Broad affected-versus-unaffected screening achieved high performance (accuracy 0.893; AUC 0.956), although sensitivity analyses revealed substantial influences of cohort composition and demographic differences. Fine-grained syndrome diagnosis remained more challenging, with limited top-1 accuracy despite meaningful class-wise discrimination. Combining repeated scans improved patient-level diagnosis beyond differences in patient weighting. Prototype inference further improved diagnostic prioritization, placing the correct syndrome among the top five candidates for about 65\% of held-out patients across 49 diagnoses. Patient-to-patient retrieval identified phenotypically similar confirmed cases, while useful syndrome prototypes could be constructed from relatively few reference patients. Anatomical region-occlusion analysis revealed syndrome-dependent patterns of facial importance broadly consistent with clinically recognized craniofacial features, although overlapping facial phenotypes contributed to diagnostic confusion. Overall, these findings support the potential of 3D facial representations for diagnostic candidate prioritization and retrieval, while emphasizing the influence of cohort composition on screening performance and the challenges of distinguishing phenotypically similar genetic syndromes.

Authors

Institutions

Publication Details

Journal
bioRxiv (Cold Spring Harbor Laboratory)
Published
2026-11-14
DOI
https://doi.org/10.64898/2026.11.12.766227
Primary Topic
Face recognition and analysis
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

From Screening to Differential Diagnosis: Benchmarking 3D Facial Representations Across Craniofacial Conditions

Mian Umair Ahsan, Eric Chien-Wei Liao, Justin L. Cotney, Tingting Chen et al.
bioRxiv (Cold Spring Harbor Laboratory)
Face recognition and analysis
preprint

From Screening to Differential Diagnosis: Benchmarking 3D Facial Representations Across Craniofacial Conditions

Mian Umair Ahsan, Eric Chien-Wei Liao, Justin L. Cotney, Tingting Chen, Kai Wang
preprint en

Abstract

Rare genetic disorders often present with highly heterogeneous physical features: some have distinctive craniofacial characteristics, whereas others produce subtler or more variable facial differences, making phenotype-based diagnosis challenging. Three-dimensional (3D) facial scanning is a noninvasive approach that captures facial shape and surface geometry, providing quantitative information on craniofacial morphology beyond conventional 2D photographs. Here, we investigated how well 3D facial representations support clinical tasks ranging from broad screening to fine-grained differential diagnosis without landmark registration, dense correspondence, or template fitting. We benchmarked four 3D models, including PointNet, PointNet++, DGCNN, and a geometric baseline, using 26,990 facial scans from 18,785 subjects across four prediction tasks. Across models, performance declined as greater diagnostic specificity was required. Broad affected-versus-unaffected screening achieved high performance (accuracy 0.893; AUC 0.956), although sensitivity analyses revealed substantial influences of cohort composition and demographic differences. Fine-grained syndrome diagnosis remained more challenging, with limited top-1 accuracy despite meaningful class-wise discrimination. Combining repeated scans improved patient-level diagnosis beyond differences in patient weighting. Prototype inference further improved diagnostic prioritization, placing the correct syndrome among the top five candidates for about 65\% of held-out patients across 49 diagnoses. Patient-to-patient retrieval identified phenotypically similar confirmed cases, while useful syndrome prototypes could be constructed from relatively few reference patients. Anatomical region-occlusion analysis revealed syndrome-dependent patterns of facial importance broadly consistent with clinically recognized craniofacial features, although overlapping facial phenotypes contributed to diagnostic confusion. Overall, these findings support the potential of 3D facial representations for diagnostic candidate prioritization and retrieval, while emphasizing the influence of cohort composition on screening performance and the challenges of distinguishing phenotypically similar genetic syndromes.

bioRxiv (Cold Spring Harbor Laboratory)
Children's Hospital of Philadelphia (US), University of Pennsylvania (US)
Face recognition and analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon