4D-FED-GNN++

4D-FED-GNN++ predicts temporal evolution of brain connectivity networks using a federated graph neural network framework that integrates decentralized longitudinal neuroimaging datasets.

Key Features:

• Federated Learning Framework: Implements a federated learning architecture that enables collaborative model training across multiple institutions while maintaining decentralized data storage.
• Graph Neural Network Modeling: Uses a graph neural network (GNN) architecture to represent and analyze brain connectivity networks and their temporal changes.
• Non-IID Longitudinal Data Handling: Supports training on non-independent and identically distributed (non-IID) longitudinal datasets collected at heterogeneous acquisition timepoints.
• Dual-Mode Prediction Strategy: Operates as a graph self-encoder when only baseline observations are available and as both a graph generator and self-encoder when follow-up connectivity data are present.
• Dual Federation Strategy: Combines layer-wise weight aggregation and randomized pairwise weight exchange among participating institutions to improve learning robustness.
• Incomplete Timepoint Optimization: The 4D-FED-GNN++ variant orders local institutions based on incomplete sequential patterns to improve federated training with missing or irregular longitudinal data.

Scientific Applications:

• Brain Connectivity Evolution Modeling: Predicts temporal changes in structural or functional brain connectivity networks.
• Neurological Disease Research: Supports analysis of connectivity alterations associated with neurological disease progression.
• Federated Neuroimaging Studies: Enables multi-institutional modeling of brain network dynamics without direct sharing of sensitive neuroimaging data.

Methodology:

4D-FED-GNN++ trains graph neural networks on decentralized longitudinal brain connectivity datasets using federated learning with layer-wise weight aggregation and randomized pairwise weight exchange, operating as a graph self-encoder or combined graph generator and self-encoder depending on data availability.