SplitAVG

SplitAVG mitigates data heterogeneity in federated learning for medical imaging by splitting neural network architectures and concatenating feature maps to enable collaborative model training across institutions without sharing patient-level data.

Key Features:

• Heterogeneity-Aware Approach: Manages variability in data distributions across multiple institutions to reduce performance degradation in federated learning.
• Simplified Methodology: Employs network splitting and feature map concatenation to create an unbiased estimator of the target data distribution, reducing reliance on complex heuristics and extensive hyperparameter tuning.
• Performance Under Heterogeneity: Maintains high performance under heterogeneous conditions, achieving 96.2% of the baseline accuracy and 110.4% of the mean absolute error in tests involving a diabetic retinopathy binary classification dataset and a bone age prediction dataset.
• Adaptability and Generalization: Adapts to various base convolutional neural networks (CNNs) and generalizes across different medical imaging tasks.

Scientific Applications:

• Federated Medical Imaging Training: Enables privacy-preserving collaborative model training across institutions without direct patient-level data sharing.
• Multi-institutional Studies: Supports analysis of heterogeneous datasets from large-scale patient cohorts and diverse demographic groups.

Methodology:

Splitting the neural network architecture and concatenating feature maps from different institutions to form an unbiased estimator that aligns the learning process with the target data distribution.