Fusion Learning from Dynamic Functional Connectivity: Combining the Amplitude and Phase of fMRI Signals to Identify Brain Disorders

A team of researchers has introduced a novel AI framework called Multi-Scale Fusion Learning (MSFL) that significantly improves the detection of brain disorders by analyzing functional magnetic resonance imaging (fMRI) data in a more comprehensive way. Traditional methods for studying dynamic functional connectivity (dFC) in the brain, like the sliding window correlation (SWC) technique, primarily focus on the amplitude of neural signals—essentially how strong the signal is over time. The MSFL framework innovates by also incorporating phase synchronization (PS), which measures the timing and coherence of signal oscillations between different brain regions. This dual approach captures a richer picture of brain network dynamics.

The researchers rigorously tested MSFL's efficacy on two major public datasets: the ABIDE I dataset for autism spectrum disorder (ASD) and the REST-meta-MDD dataset for major depressive disorder (MDD). The results demonstrated that MSFL's integrated analysis of both amplitude correlation and phase coherence led to significantly better classification performance than models using either feature type alone or other existing comparative methods. To explain the model's decisions and validate the contribution of each data type, the team used the SHAP (SHapley Additive exPlanations) framework. This analysis confirmed that both the SWC-derived amplitude features and the PS-derived phase features were important for the model's accurate identification of brain disorders, providing a clearer, more interpretable path to diagnosis.