3D Facial Pattern Analysis for Autism Using Geodesic Distances

Thursday, May 15, 2014
Atrium Ballroom (Marriott Marquis Atlanta)
T. Obafemi-Ajayi1, J. H. Miles2, W. Qi1, N. Takahashi2, K. Aldridge3, Y. Duan1 and H. Ying4, (1)University of Missouri, Columbia, MO, (2)Thompson Center for Autism and Neurodevelopmental Disorders, University of Missouri, Columbia, MO, (3)University of Missouri School of Medicine, Columbia, MO, (4)Nanyang Technological University, Singapore, Singapore

The facial analysis is of great importance as it is related to brain development: “face mirrors brain”. Recent studies suggest statistically significant differences in facial morphology for the various subgroups of children using 3D facial imaging data based on Euclidean distances. In this study, we apply 3D facial geodesic (surface) distance features as well as robust clustering analysis to identify etiologically discrete ASD subgroups.


To identify clinically meaningful clusters within ASD and the facial features that discriminate these clusters using 3D geodesic (surface) distance features. We seek to identify meaningful heterogeneity within the ASD dataset in comparison to the traditional autism subtypes (Autistic Disorder, Asperger Syndrome and PDD-NOS).


We used the 3dMD software to obtain 3D coordinate data for a set of 19 anthropometric facial landmarks of our study sample of 62 ASD boys. All the boys were Caucasian between the ages of 8 and 12. To identify the potential subgroups, we applied different clustering techniques: Expectation Maximization, K-means, Partitioning around Medoids, and Self-Organizing Feature Map. We selected the optimal cluster configuration based on the cluster validation evaluation. The cluster results were validated clinically.


We identified 3 clusters with distinct geodesic facial features as well as meaningful clinical and behavioral traits. The ASD subtypes (Autistic Disorder, Asperger Syndrome and PDD-NOS) were not evenly distributed within the 3 Clusters.  Cluster 1 consists of mainly Autistic Disorder subjects (78.6%). Cluster 2 subjects consist of half Autistic Disorder (50%) and almost half Asperger Syndrome (44.4%) as well as one PDD boy. The Cluster 3 group is a mixed bouquet of diagnoses: (46.7% Autistic Disorder: 33.3% Asperger Syndrome: 20.0% PDD-NOS). PDD-NOS, the least precise autism diagnosis clinically, appears to be present most in the Cluster 3 group. The Cluster 1 group seemed to be the most compact and separate group facially, clinically and behaviorally. It consisted mainly of severe Autistic Disorder boys.


There is a facial difference among autistic subjects. These facial differences show a spectrum of scale among the Autism diagnosis – Asperger Syndrome vs. Autistic Disorder. This implies that there are physical phenotypes related to Autism not just behavioral traits.