Classifying ASD based on time-series fMRI using spatial-temporal transformer.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Autism spectrum disorder (ASD) prevalence is rising globally, necessitating efficient diagnostic tools.
• Current ASD diagnosis relies on subjective, time-consuming clinical observations.
• Functional magnetic resonance imaging (fMRI) shows promise for identifying objective biomarkers for ASD.

Purpose of the Study:

• To develop and evaluate a deep learning framework, the spatial-temporal Transformer (ST-Transformer), for distinguishing individuals with ASD from typical controls using fMRI data.
• To address data imbalance issues in real-world ASD datasets for improved subtype diagnosis.

Main Methods:

• Developed a novel deep learning framework: spatial-temporal Transformer (ST-Transformer).
• Proposed a linear spatial-temporal multi-headed attention unit to capture fMRI data's spatial and temporal features.
• Implemented a Gaussian Generative Adversarial Network (GAN)-based data balancing method to handle dataset imbalances.

Main Results:

• The ST-Transformer achieved robust diagnostic accuracies of 71.0% on the ABIDE I dataset and 70.6% on the ABIDE II dataset.
• Demonstrated competitive performance compared to existing state-of-the-art methods in ASD diagnosis.
• Successfully utilized fMRI data for objective ASD identification.

Conclusions:

• The ST-Transformer framework shows significant potential as an objective tool for ASD diagnosis using fMRI.
• The proposed methods effectively address challenges in ASD dataset analysis, including data imbalance.
• This deep learning approach offers a promising advancement in the field of neuroimaging-based ASD detection.