Diffusion transformer-augmented fMRI functional connectivity for enhanced autism spectrum disorder diagnosis

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Functional magnetic resonance imaging (fMRI) analysis relies on understanding brain networks and functional connectivity.
• High costs of fMRI data acquisition limit the performance of recognition models.
• Data scarcity is a significant challenge in brain imaging research.

Purpose of the Study:

• To address the challenge of limited fMRI data using generative diffusion models for data augmentation.
• To introduce Brain-Net-Diffusion, a novel transformer-based latent diffusion model.
• To evaluate the impact of generated functional connectivity on fMRI classification tasks.

Main Methods:

• Developed Brain-Net-Diffusion, a transformer-based latent diffusion model.
• Generated realistic functional connectivity patterns for augmenting fMRI datasets.
• Assessed the model's performance on downstream classification tasks, including autism spectrum disorder detection.

Main Results:

• Brain-Net-Diffusion successfully generated functional connectivity patterns that mimic real fMRI data.
• Data augmentation with Brain-Net-Diffusion significantly improved classification performance.
• Autism spectrum disorder classification accuracy increased by 4.3% compared to no augmentation, outperforming other methods by 1.3%–2.2%.

Conclusions:

• Diffusion models offer an effective solution for fMRI data augmentation, overcoming data scarcity.
• Brain-Net-Diffusion enhances the robustness of functional connectivity analysis.
• The study provides a valuable tool for advancing brain imaging research, with code publicly available.