Age-Specific Diagnostic Classification of ASD Using Deep Learning Approaches

Abstract

Autism Spectrum Disorder (ASD) is a highly heterogeneous condition, due to high variance in its etiology, comorbidity, pathogenesis, severity, genetics, and brain functional connectivity (FC). This makes it devoid of any robust universal biomarker. This study aims to analyze the role of age and multivariate patterns in brain FC and their accountability in diagnosing ASD by deep learning algorithms. We utilized functional magnetic resonance imaging data of three age groups (6 to 11, 11 to 18, and 6 to 18 years), available with public databases ABIDE-I and ABIDE-II, to discriminate between ASD and typically developing. The blood-oxygen-level dependent time series were extracted using the Gordon's, Harvard Oxford and Diedrichsen's atlases, over 236 regions of interest, as 236x236 sized FC matrices for each participant, with Pearson correlations. The feature sets, in the form of FC heat maps were computed with respect to each age group and were fed to a convolutional neural network, such as MobileNetV2 and DenseNet201 to build age-specific diagnostic models. The results revealed that DenseNet201 was able to adapt and extract better features from the heat maps, and hence returned better accuracy scores. The age-specific dataset, with participants of ages 6 to 11 years, performed best, followed by 11 to 18 years and 6 to 18 years, with accuracy scores of 72.19%, 71.88%, and 69.74% respectively, when tested using the DenseNet201. Our results suggest that age-specific diagnostic models are able to counter heterogeneity present in ASD, and that enables better discrimination.