Identification of Autism Subtypes Based on Wavelet Coherence of BOLD FMRI Signals Using Convolutional Neural Network.

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Resting-state functional magnetic resonance imaging (rs-fMRI) functional connectivity (FC) is crucial for autism spectrum disorder (ASD) classification.
• Existing models using FC patterns achieve less than 80% accuracy and lack multi-site generalizability.
• There is a need for models that can identify specific ASD subtypes.

Purpose of the Study:

• To develop an automated system for identifying ASD subtypes using convolutional neural networks (CNNs).
• To utilize dynamic FC patterns derived from wavelet coherence scalograms as input features.
• To evaluate the model's performance in both binary (ASD vs. Normal Control) and multi-class (ASD subtypes vs. Normal Control) classifications.

Main Methods:

• rs-fMRI data from 144 individuals across 8 sites, labeled into three ASD subtypes (ASD, APD, PDD-NOS) and Normal Control (NC).
• Identification of the top-ranked brain node (putamen_R) using power spectral density (PSD) analysis.
• Computation of wavelet coherence scalograms between the top node and other nodes to represent dynamic FC.
• Development and testing of CNN models using cross-validation and leave-one-out techniques.

Main Results:

• Binary classification (ASD vs. NC) achieved 89.8% accuracy.
• Multi-class classification (ASD vs. APD vs. PDD-NOS vs. NC) yielded a 82.1% macro-average accuracy.
• The wavelet coherence technique effectively represented dynamic FC patterns.

Conclusions:

• The proposed CNN model using dynamic FC from wavelet coherence shows significant potential for automated ASD subtype identification.
• This method demonstrates improved accuracy over existing approaches and suggests generalizability.
• The wavelet coherence technique offers a promising tool for analyzing brain connectivity in neuropsychiatric disorders.