Conditional Variational Autoencoder for Functional Connectivity Analysis of Autism Spectrum Disorder Functional...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Autism Spectrum Disorder Research

Background:

• Generative models, specifically Variational Autoencoders (VAEs), are emerging tools for detecting abnormal patterns in brain imaging data.
• These models learn typical brain patterns to identify neurodivergent states by measuring deviations in generated or reconstructed images.
• Functional Connectivity (FC) analysis using fMRI data is crucial for understanding brain region interconnectivity.

Purpose of the Study:

• To evaluate the efficacy of different VAE architectures for Functional Connectivity (FC) analysis in Autism Spectrum Disorder (ASD).
• To investigate whether incorporating phenotypic data enhances VAE performance for FC analysis in ASD.
• To identify atypical brain region interconnectivity in individuals with ASD.

Main Methods:

• Comparison of multiple VAE architectures: Conditional VAE, Recurrent VAE, and a hybrid CNN-RNN VAE.
• Application of VAEs to functional Magnetic Resonance Imaging (fMRI) data from individuals with ASD.
• Evaluation of VAE performance with and without the inclusion of phenotypic data.

Main Results:

• Convolutional Neural Network (CNN)-based VAE architectures demonstrated superior effectiveness for FC analysis compared to other evaluated models.
• The study identified specific atypical interconnectivity patterns in the brain associated with ASD.
• Incorporating phenotypic data showed potential for improving VAE performance in FC analysis for ASD.

Conclusions:

• CNN-based VAEs are highly effective for detecting atypical functional connectivity in Autism Spectrum Disorder (ASD).
• This approach offers a promising method for uncovering neurodivergent brain patterns.
• Further research into integrating phenotypic data could refine VAE-based neuroimaging analysis for ASD.