Researchers at InfoLab, Sungkyunkwan University, led by Professor Tamer Abuhmed, have introduced the 4D fMRI CrossFormer (4DfCF), a vision transformer architecture tailored for four-dimensional functional magnetic resonance imaging. Unlike conventional approaches that flatten or reduce fMRI data, this model treats the full 4D structure—three spatial dimensions plus time—as a unified input, allowing it to capture dynamic brain activity more naturally.
The architecture is designed to be computationally efficient, addressing a common bottleneck in high-dimensional medical imaging. By using cross-attention mechanisms across spatial and temporal dimensions, the model avoids the heavy memory and processing demands of earlier transformer-based methods, making it more practical for clinical settings.
A key advantage is transparency. The model's attention maps can be visualized to show which brain regions and time points contribute most to a screening decision, offering clinicians a clearer rationale than typical black-box AI systems. The team reports that this approach maintains competitive accuracy while improving interpretability and reducing computational overhead, a step toward more trustworthy AI-assisted diagnosis of brain disorders.