Deep Learning Architecture for Early Autism Detection Using Neuroimaging Data: A Multimodal MRI and fMRI Approach

Posted: Jul 10, 2018

• Hammad Khan
  Punjab College
• Michael Johnson
  University of Illinois Urbana-Champaign
• Emily Smith
  University of Illinois Urbana-Champaign

Abstract

Autism Spectrum Disorder (ASD) represents a complex neurodevelopmental
condition characterized by challenges in social interaction, communication, and
restricted or repetitive behaviors. Early detection of ASD is crucial for timely in
tervention and improved long-term outcomes. This research presents a novel deep
learning architecture specifically designed for early autism detection using mul
timodal neuroimaging data. Our framework integrates both structural Magnetic
Resonance Imaging (sMRI) and functional Magnetic Resonance Imaging (fMRI)
to capture complementary neurobiological signatures of ASD. The proposed model
employs a dual-pathway convolutional neural network for processing structural fea
tures from sMRI and a recurrent neural network with attention mechanisms for
analyzing functional connectivity dynamics from fMRI. The multimodal features
are subsequently fused through a carefully designed integration module. We eval
uated our approach on the extensively used ABIDE I and II datasets, comprising
over 2,000 subjects from multiple imaging sites. Our model achieved a classification
accuracy of 92.7%, sensitivity of 91.3%, and specificity of 93.8%, significantly out
performing existing single-modality approaches and traditional machine learning
methods. The attention mechanisms within our architecture provide interpretable
insights by highlighting brain regions most discriminative for ASD classification,
particularly in the default mode network, salience network, and frontotemporal
pathways. This research establishes a robust foundation for computer-aided early
autism diagnosis and offers promising directions for clinical translation.