Diagnosing Autism Spectrum Disorder from Brain Resting-State Functional Connectivity Patterns Using a Deep Neural...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Resting-state functional magnetic resonance imaging (rs-fMRI) reveals whole-brain functional connectivity (FC) patterns relevant to neuropsychiatric conditions like autism spectrum disorder (ASD).
• Aberrant FCs in ASD are widespread, affecting multiple brain regions, necessitating advanced analytical techniques for accurate classification.
• Deep neural networks (DNNs) show promise in extracting complex information from high-dimensional data, improving classification accuracy in neurological studies.

Purpose of the Study:

• To develop and evaluate a novel deep neural network with a feature selection method (DNN-FS) for classifying autism spectrum disorder (ASD) based on whole-brain resting-state functional connectivity (FC) patterns.
• To compare the performance of the DNN-FS approach against a DNN without feature selection (DNN-woFS) and other traditional feature selection methods.
• To identify potential FC-based biomarkers for ASD using a Fisher's score-based method integrated with the DNN.

Main Methods:

• A deep neural network with a novel feature selection method (DNN-FS) was developed, utilizing sparse auto-encoders to select high-discriminating power features from whole-brain resting-state FC patterns.
• The DNN-FS model was compared with a DNN without feature selection (DNN-woFS) across various architectures (different numbers of hidden layers and nodes).
• A Fisher's score-based method was employed to identify ASD-related FCs from the DNN model, assessing their statistical significance and relationship to behavioral symptoms.

Main Results:

• The DNN-FS approach achieved a maximum classification accuracy of 86.36% with a 3-hidden-layer, 150-node architecture (3/150).
• DNN-FS consistently outperformed DNN-woFS across all tested architectures, with the most significant accuracy improvement being 9.09% for the 3/150 model.
• The method demonstrated superior performance compared to traditional feature selection techniques like two-sample t-test and elastic net, and identified 32 FCs related to ASD.

Conclusions:

• The developed DNN-FS method offers a powerful and accurate approach for classifying autism spectrum disorder (ASD) using resting-state functional connectivity (FC) data.
• Feature selection significantly enhances the performance of deep neural networks in identifying complex patterns within high-dimensional neuroimaging data.
• The identified FCs and the DNN-based biomarker discovery approach hold potential for advancing ASD diagnosis and understanding its underlying neural mechanisms.