Towards Personalized Medicine in Autism Diagnosis: Anatomical Abnormalities Analysis Using a Deep Learning Based Approach

Oral Presentation
Saturday, May 12, 2018: 3:16 PM
Jurriaanse Zaal (de Doelen ICC Rotterdam)
O. Dekhil1, M. Ghazal2, A. Shalaby3, A. Switala4, A. S. El-Baz4, A. Khalil5 and G. Barnes6, (1)Bioengineering, university of Louisville, Louisville, KY, (2)Electrical and Computer Engineering, Abu Dhabi University, Abu Dhabi, United Arab Emirates, (3)Bioengineering, University of Louisville, Louisville, KY, (4)University of Louisville, Louisville, KY, (5)Information Technology and Computer Science, Abu Dhabi University, Abu Dhabi, United Arab Emirates, (6)University of Louisville School of Medicine, Louisville, KY
Background: Background. Autism spectrum disorder (ASD) is a neurodevelopmental syndrome that affects both communication and social interaction. Although there is no well-defined cause of ASD, there are many proposed theories and hypotheses aiming to explain ASD causes; for example structural abnormalities, connectivity abnormalities and functional activation are widely studied.

Objectives: In this study data from structural MRI (sMRI) was used to identify anatomical characteristics distinguishing ASD from typically developing (TD) subjects in a demonstration of Sight software for computer-assisted diagnosis (CAD) of ASD, developed at our laboratory.

Methods: The Sight processing pipeline begins by isolating the brain from a T1-weighted MRI using Fsl Bet with post-processing to validate the result. The test brain is aligned with a library of previously segmented data, which are used to derive prior shape information to guide the segmentation of the cerebral cortex. A triangulated mesh of the cortical surface is output. Mesh vertex coordinates are re-parameterized in terms of spherical polar coordinates in order to analytically approximate the surface by spherical harmonic series. Three shape descriptors are obtained per Brodmann area: (a) truncation error as a function of order of spherical harmonic approximation, (b) distribution of mean curvature, and (c) distribution of Gaussian curvature. Features are input to a stacked autoencoder, producing higher-order representations, which are input to a second network with a softmax classification layer, trained to recognize features consistent with ASD on a local level. The higher-order representations from all areas are also concatenated for input to another classifier to obtain the overall decision. The Sight classifier was trained and tested using scans of 202 individuals downloaded from the National Database for Autism Research (NDAR), 78 of whom were diagnosed with ASD.

Results: ASD and TD subgroups both ranged in age from 7 to 17 years, and had similar proportions of male and female individuals. Using two-fold cross validation the Sight software correctly classified 78 of 78 ASD cases and 90 of 124 TD cases. Some brain regions were found to have significant influence on classification. Using supramarginal gyrus, pars triangularis, or inferior prefrontal gyrus, to the exclusion of all other regions, produced classifiers with 76%, 77%, and 79% accuracy, respectively. The most influential areas are highly correlated with functional deficits in autism, such as speech and syntax processing and emotional responses.

Conclusions: Sight is an important step towards personalized medicine, where each localized brain region is studied separately to allow better way to allocate each subject on the autism spectrum and enables better prediction to the affected brain functionality based on the affected areas, but expansions to the framework are already in progress to: (1) incorporate dwMRI- and fMRI-based descriptors into a multimodal CAD system that can better resolve ASD endophenotypes and (2) distinguish ASD from other disorders such as childhood epilepsy.

Tech demo. There will be live demonstration of Sight CAD software. All procedure steps from segmentation to report generation will be performed, and intermediate output will be available for inspection.