Automatic characterization of cerebral MRI images for the detection of autism spectrum disorders

Abstract

Autism Spectrum Disorders (ASD) are one of the most serious health problems that our generation is facing [1]. It affects around one out of every 54 children and causes issues with social interaction, communication [2] and repetitive behaviors [3]. The development of full biomarkers for neuroimaging is a crucial step in diagnosing and tailoring medical care for autism spectrum disorder [4]. Volumetric studies focused on 3D MRI texture features have shown a high capacity for detecting abnormalities and characterizing variations caused by tissue heterogeneity. Recently, it has been the interest of comprehensive studies. However, only a few studies have aimed to investigate the link between object texture and ASD. This paper suggests a framework based on geometric texture features analyzing the variations between ASD and development control (DC) subjects. Our study uses 1114 T1-weighted MRI scans from two groups of subjects: 521 individuals with ASD and 593 controls (age range: 6–64 years) [5], divided into three broad age groups. We then computed the features from automatically labeled subcortical and cortical regions and encoded them as texture features by applying seven global Riemannian geometry descriptors and eight local features of standard Harlicks quantifier functions. Significant tests were used to identify texture volumetric differences between ASD and DC subjects. The most discriminative features are selected by applying the Correlation Matrix, and these features are used to classify the two classes using an Artificial Neural Network analysis. Preliminary results indicate that in ASD subjects, all 15 structure-derived features and subcortical regions tested have significantly different distributions from DC subjects.