Explaining deep learning-based representations of resting state functional connectivity data: focusing on...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Psychiatry

Background:

• Resting state functional Magnetic Resonance Imaging (rs-fMRI) is crucial for studying brain organization in psychiatric disorders.
• High-dimensional rs-fMRI data analysis is challenging.
• Variational Autoencoders (VAEs) reduce dimensionality but lack interpretability.

Purpose of the Study:

• Develop explainable VAE models for rs-fMRI data.
• Interpret latent representations of resting state functional connectivity (rsFC).
• Test utility in autism spectrum disorder (ASD).

Main Methods:

• Trained VAEs on rs-fMRI data from 1150 participants (549 ASD, 601 controls).
• Extracted rsFC matrices using the Power atlas (264 ROIs).
• Introduced latent contribution scores to link representations to rs-fMRI measures.

Main Results:

• Both ASD and control groups share top network connectivities.
• Latent 0 was driven by the ventral attention network (VAN) in both groups.
• Significant differences in latent contribution scores were found in VAN (latent 0) and sensory/somatomotor network (latent 2) between ASD and HC groups.

Conclusions:

• Latent contribution scores enable interpretation of VAE-identified nonlinear patterns.
• Explainable VAEs enhance understanding of ASD neural mechanisms.
• This approach facilitates deeper insights into brain function in psychiatric disorders.