Reproduce “Structural Connectome Harmonization Using Deep Learning: The Strength of Graph Neural Networks” on OpenResearch

About this paper

Jagruti Patel, Thomas A. W. Bolton, Mikkel Schöttner, Anjali Tarun, Sebastien Tourbier, Yasser Alemàn-Gòmez, Jonas Richiardi, Patric Hagmann

Researchers from Lausanne University Hospital and University of Lausanne developed a deep learning framework for harmonizing structural connectomes from multi-site neuroimaging studies, focusing on preserving individual-specific and topological brain network properties. Their Graph Autoencoder (GAE) model demonstrated superior individuality retention and topological preservation without needing detailed acquisition metadata, outperforming other deep learning models and a linear regression baseline in practical multi-site scenarios.

Key insights

• Graph Neural Networks (GAE) are particularly effective for SC harmonization in scenarios where explicit acquisition metadata is unavailable, demonstrating superior preservation of subject individuality and topological properties compared to other deep learning models.
• While linear regression (LR) excels in edge-level accuracy and some topological measures when acquisition parameters are explicitly known, its practical applicability is limited by the frequent unavailability of such detailed metadata.
• The adversarial training with a site-classifier successfully disentangles site-specific biases from subject-specific biological information within the latent space, with GAE achieving the best separation for robust subject-level clustering.

Summary by alphaXiv.