Multiple-site diffusion MRI tractography analysis using federated learning for brain disease classification

Motivation: Deep learning shows great potential in diffusion MRI (dMRI) tractography for clinical and disease classification, but it is challenged by the limited size of study samples. Federated learning (FL) offers an effective way to address this, by decentralized model training on local datasets. Goal(s): We propose the first FL framework to enable multiple-site dMRI tractography analysis for disease classification. Approach: A novel and effective site-weighting strategy is presented to address the data distribution differences across sites. Furthermore, we design a model interpretation module to pinpoint the discriminative brain regions between the groups. Results: Our approach achieves 80.1% accuracy for Autism Spectrum Disorder classification. Impact: This study presents the first deep federated learning framework to enable multiple-site dMRI tractography analysis for disease classification. The novel and site-weighting strategy can effectively accommodate data distribution differences across sites by demonstrating on Autism Spectrum Disorder classification.