Computational Modeling of Facial Response for Detecting Differential Traits in Autism Spectrum Disorders

This dissertation proposes novel computational modeling and computer vision methods for the analysis and discovery of differential traits in subjects with Autism Spectrum Disorders (ASD) using video and three-dimensional (3D) images of face and facial expressions. ASD is a neurodevelopmental disorder that impairs an individual’s nonverbal communication skills. This work studies ASD from the pathophysiology of facial expressions which may manifest atypical responses in the face. State-of-the-art psychophysical studies mostly employ na¨ıve human raters to visually score atypical facial responses of individuals with ASD, which may be subjective, tedious, and error prone. A few quantitative studies use intrusive sensors on the face of the subjects with ASD, which in turn, may inhibit or bias the natural facial responses of these subjects. This dissertation proposes non-intrusive computer vision methods to alleviate these limitations in the investigation for differential traits from the spontaneous facial responses of individuals with ASD. Two IRB-approved psychophysical studies are performed involving two groups of age-matched subjects: one for subjects diagnosed with ASD and the other for subjects who are typically-developing (TD). The facial responses of the subjects are computed from their facial images using the proposed computational models and then statistically analyzed to infer about the differential traits for the group with ASD. A novel computational model is proposed to represent the large volume of 3D facial data in a small pose-invariant Frenet frame-based feature space. The inherent pose-invariant property of the proposed features alleviates the need for an expensive 3D face registration in the pre-processing step. The proposed modeling framework is not only computationally efficient but also offers competitive performance in 3D face and facial expression recognition tasks when compared with that of the state-ofthe-art methods. This computational model is applied in the first experiment to quantify subtle facial muscle response from the geometry of 3D facial data. Results show a statistically significant asymmetry in specific pair of facial muscle activation (p\u3c0.05) for the group with ASD, which suggests the presence of a psychophysical trait (also known as an ’oddity’) in the facial expressions. For the first time in the ASD literature, the facial action coding system (FACS) is employed to classify the spontaneous facial responses based on facial action units (FAUs). Statistical analyses reveal significantly (p\u3c0.01) higher prevalence of smile expression (FAU 12) for the ASD group when compared with the TD group. The high prevalence of smile has co-occurred with significantly averted gaze (p\u3c0.05) in the group with ASD, which is indicative of an impaired reciprocal communication. The metric associated with incongruent facial and visual responses suggests a behavioral biomarker for ASD. The second experiment shows a higher prevalence of mouth frown (FAU 15) and significantly lower correlations between the activation of several FAU pairs (p\u3c0.05) in the group with ASD when compared with the TD group. The proposed computational modeling in this dissertation offers promising biomarkers, which may aid in early detection of subtle ASD-related traits, and thus enable an effective intervention strategy in the future