HSGO: Harmonized Swarm Learning With Guided Optimization for Multi-Center sMRI Classification of Alzheimer's Disease

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Informatics

Background:

• Developing robust Alzheimer's Disease (AD) classification models requires extensive multi-center data, but data aggregation raises privacy concerns.
• Existing Federated Learning (FL) and Swarm Learning (SL) methods face performance limitations due to data heterogeneity and class imbalance across centers.

Purpose of the Study:

• To propose a novel Harmonized Swarm Learning framework with Guided Optimization (HSGO) for privacy-preserving, enhanced multi-center collaboration in AD classification.
• To improve the robustness and performance of generic and personalized AD classification models trained across diverse datasets.

Main Methods:

• Implemented a class-balanced loss function for training a robust generic model within the HSGO framework.
• Utilized guided optimization to align personalized models with the generic model, avoiding additional feature extraction.
• Developed a dynamic feature similarity storage mechanism to support personalized model training.

Main Results:

• HSGO demonstrated competitive performance compared to five baseline methods across two multi-center data partitioning scenarios.
• Layer-wise Relevance Propagation (LRP) analysis suggested HSGO's potential in identifying key brain regions for AD by integrating local and global features.

Conclusions:

• HSGO offers an effective solution for privacy-preserving multi-center collaboration in AD classification.
• The framework shows promise in enhancing model performance and potentially aiding in the identification of critical neuroimaging biomarkers for Alzheimer's Disease.