Skip to main content
. 2026 Oct 7;20:1869687. doi: 10.3389/fninf.2026.1869687

Figure 2.

Flowchart detailing an fMRI preprocessing pipeline with five color-coded steps: raw 4D fMRI data, structural preprocessing (skull stripping, tissue segmentation, normalization), functional preprocessing (motion correction, functional normalization, smoothing), denoising (regression, filtering), and feature engineering (voxelization, ROI extraction, functional connectivity matrix), concluding with transformer token sequence input.

Processing workflow from raw 4D fMRI data to Transformer input tokens: (1) Raw 4D fMRI data; (2) Structural image preprocessing and normalization; (3) Head motion correction and smoothing of functional images; (4) Removal of physiological noise and signal drift through nuisance regression and temporal filtering; (5) Feature engineering to construct three types of features, voxel patches, ROI time series, and functional connectivity matrices, and conversion into token sequences.