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

3D facial models can prioritize rare craniofacial syndromes without landmark registration, with the correct syndrome in the top five for about 65% of held-out patients across 49 diagnoses.

• Why it matters: Phenotype-based diagnosis is hard because rare genetic disorders show heterogeneous and sometimes subtle facial differences, so clinicians need quantitative tools that work from raw 3D scans.
• What they did: The study benchmarked PointNet, PointNet++, DGCNN, and a geometric baseline on 26,990 scans from 18,785 subjects across four tasks, including affected-versus-unaffected screening and fine-grained syndrome diagnosis.
• The result: Screening reached accuracy 0.893 and AUC 0.956, but performance dropped as specificity increased; prototype inference and repeated scans improved patient-level diagnosis and retrieval of phenotypically similar cases.

AI summary of the abstract · may contain errors · not medical advice · read the paper.