Research & Papers

Researchers unveil FedDOSE to improve brain disorder detection

New federated learning framework detects autism and ADHD with 15% higher accuracy in multi-site fMRI scans.

Deep Dive

Researchers from Nanyang Technological University have introduced **FedDOSE**, a federated learning framework designed to improve brain disorder detection by addressing statistical heterogeneity in multi-site functional MRI (fMRI) datasets. Traditional federated learning approaches often struggle with site-specific biases, where data collected from different imaging centers introduces variability that degrades model performance. FedDOSE tackles this by decomposing site effects using a **Modularity-Guided Tucker Decomposition** block, which encodes high-dimensional dynamic functional connectivity (dFC) tensors to capture modular-level spatio-temporal patterns efficiently.

The framework enhances cross-site generalization by generating **class-specific prototypes** and aligning them globally using **Optimal Transport (OT) barycenter formulation** and **Procrustes analysis**. In experiments diagnosing Autism Spectrum Disorder (ASD) and Attention-Deficit Hyperactivity Disorder (ADHD) across three datasets (ABIDE-I, ABIDE-II, ADHD-200), FedDOSE achieved **15% higher accuracy** than existing state-of-the-art methods. This breakthrough enables more reliable and privacy-preserving multi-site neuroimaging studies.

Key Points
  • FedDOSE improves ASD and ADHD detection accuracy by 15% in multi-site fMRI datasets (ABIDE-I, ABIDE-II, ADHD-200).
  • The framework uses Modularity-Guided Tucker Decomposition and Optimal Transport alignment to handle site-specific biases.
  • FedDOSE enables privacy-preserving collaborative neuroimaging research across multiple institutions.

Why It Matters

Enables more accurate and scalable brain disorder diagnosis while preserving patient privacy across global research consortia.

📬 Get the top 10 AI stories daily