DSFM: New AI framework generates realistic fMRI data for brain disorder detection
Dual-spectral flow matching creates synthetic brain scans that rival real ones.
Functional MRI data is crucial for studying brain disorders, but acquiring high-quality samples is expensive and time-consuming. A new framework called Dual-Spectral Flow Matching (DSFM), accepted at ICLR 2026, tackles this by generating synthetic fMRI time series that preserve the complex, non-stationary dynamics of real BOLD signals. The approach first converts raw signals into a wavelet decomposition map to capture multi-scale temporal variations, then projects them into discrete cosine transform (DCT) space to exploit energy compaction in low-frequency components. A spectral flow matching model then learns to generate class-conditioned cosine-frequency representations, which are inverted back into time-domain signals that are physiologically plausible.
DSFM outperforms existing generative models on downstream tasks such as brain network classification for disorder identification. By imposing structured frequency priors, the framework ensures the synthetic data retains key biomarkers from real scans. The code is open-source, making it accessible for researchers to augment limited fMRI datasets. This could accelerate the development of AI diagnostics for conditions like Alzheimer's, schizophrenia, and epilepsy, where large labeled datasets are scarce. DSFM represents a significant step toward reliable synthetic medical imaging for clinical AI.
- DSFM cascades discrete wavelet and cosine transforms to capture non-stationary BOLD signal dynamics.
- The spectral flow matching model generates class-conditioned frequency representations for targeted brain disorder identification.
- Accepted at ICLR 2026 with open-source code, enabling reproducible augmentation of fMRI datasets.
Why It Matters
Synthetic fMRI data could lower barriers for training AI models on rare brain disorders, improving diagnostic accuracy.