2D Facial Pattern Analysis for Autism

Saturday, May 19, 2012
Sheraton Hall (Sheraton Centre Toronto)
11:00 AM
T. Obafemi-Ajayi1, B. Morago1, J. Wilson1, T. N. Takahashi2, K. Aldridge2,3, J. H. Miles4 and Y. Duan1,2, (1)Computer Science Department, University of Missouri, Columbia, MO, (2)Thompson Center for Autism and Neurodevelopmental Disorders, Columbia, MO, (3)Pathology & Anatomical Sciences, University of Missouri, Columbia, MO, (4)Thompson Center for Autism and Neurodevelopmental Disorders, University of Missouri, Columbia, MO

Recent studies suggest that differences in facial morphology in children with autism spectrum disorder (ASD) compared to typically developing children (control) exist and may reflect alterations in embryologic brain development.  As ASD can present a wide range of symptoms, the same variations of facial morphology may help pinpoint differing forms of this disorder.  Experiments run on 3D facial images indicate statistically significant differences in facial morphology for various ASD subgroups. In this study, we investigate whether similar findings can be observed using 2D facial images, as they are more readily available and easier to capture.  Our study includes tests for distinguishing between the ASD and control groups as well as for identifying subgroups within the ASD group.


The goals are to define discriminant facial phenotypes directly from the 2D facial photographs in order to aid in the early diagnosis of ASD, as well as to identify meaningful subgroups for clinical and genetic study.


We defined facial features based on anthropometric landmark points which were extracted from the 2D facial photos.  From these landmarks, we calculated feature distances on each face by computing the Euclidean distance between pairs of landmarks.  Our study sample consisted of 172 children with autism and 54 typically developing children.  To identify potentially strong features for distinguishing between the groups, we first performed Principle Component Analysis on varying subsets of the feature distances.  This step allowed us to see which feature distances account for large variations within the dataset. The data set was then separated into groups by using Expectation–Maximization clustering. The obtained clusters were validated using Adaboost ensemble classification. The discriminant facial features from this step were selected by evaluating the ranked significance of each feature using the Info-Gain attribute evaluator and Forward Feature selection.


First, we identified significant differences in 2D facial morphology in children with ASD compared to typically developing children. Second, we observed five facial feature distances that may be highly discriminant in identifying several subgroups of ASD children from typically developing children.


2D facial images contain clinically-meaningful facial phenotype information capable of distinguishing ASD children from typically developing children as has been described previously by 3D studies.

| More