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

Functional magnetic resonance imaging (fMRI) has emerged as a transformative tool in analyzing and understanding brain diseases. It is a challenge to learn effective features from the high-dimensional fMRI. Most studies have focused on extracting connectivity-based features for disease analysis. However, they heavily rely on the software toolboxes to construct connectivity-based features, which may suffer from large errors because of different manual parameter settings and thus lead to bad performance in brain disorder analysis.

A novel brain denoiser model is proposed to transform four-dimensional fMRI (4D fMRI) into a brain network in a unified framework for brain disease analysis.

By introducing anatomical knowledge, the proposed model first reduces the 4D fMRI into a 2D coarse region-of-interest(ROI)-based time series and then diffuses it into noisy status by gradually adding Gaussian noise. Moreover, the coarse-to-fine transformer refinement is designed to capture multi-scale temporal dynamics and iteratively remove unrelated multi-frequency noise. Besides, the low-frequency preservation module is devised to enhance the effective signal at low frequencies during the denoising process. This can improve the signal-to-noise ratio at each timestep, which ensures accurate restoration of ROI time series and improves the performance of brain network construction.

We evaluate the performance of the Brain Denoiser on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset and the Autism Brain Imaging Data Exchange (ABIDE) dataset, demonstrating its ability to effectively suppress noise while preserving the underlying neural signals. Comparative analyses with related competing methods demonstrate the superiority of the proposed model.

Generally, the proposed model presents a robust and innovative solution for brain network generation, paving the way for efficient analysis of brain disease.