Spatiotemporal coarse-to-fine diffusion model for automatic brain network generation.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biomedical Engineering

Background:

• Functional magnetic resonance imaging (fMRI) is vital for brain disease research but faces challenges with high-dimensional data.
• Current methods often rely on connectivity features, which can be prone to errors due to manual parameter settings in software toolboxes.
• This can negatively impact the accuracy of brain disorder analysis.

Purpose of the Study:

• To introduce a novel brain denoiser model for transforming 4D fMRI data into brain networks within a unified framework.
• To enhance the analysis and understanding of brain diseases using improved fMRI data processing.

Main Methods:

• The model integrates anatomical knowledge to reduce 4D fMRI to 2D ROI-based time series.
• A coarse-to-fine transformer refinement captures multi-scale temporal dynamics and removes multi-frequency noise.
• A low-frequency preservation module enhances effective signals, improving signal-to-noise ratio and ROI time series restoration.

Main Results:

• The Brain Denoiser was evaluated on Alzheimer's Disease Neuroimaging Initiative (ADNI) and Autism Brain Imaging Data Exchange (ABIDE) datasets.
• The model demonstrated effective noise suppression while preserving underlying neural signals.
• Comparative analyses confirmed the proposed model's superiority over competing methods.

Conclusions:

• The proposed model offers a robust and innovative solution for generating brain networks from fMRI data.
• This approach facilitates more efficient and accurate analysis of brain diseases.
• It represents a significant advancement in neuroimaging analysis for clinical applications.