Fed-BioMed

Fed-BioMed enables federated learning on decentralized medical data to train and aggregate machine learning models for privacy-preserving analysis of partially labeled clinical imaging datasets.

Key Features:

• Federated learning framework: Implements federated learning to train models across multiple institutions on decentralized health data without centralizing patient-level data.
• Partial-label handling: Supports scenarios with partially labeled datasets where only some sites have expert annotations and others contain large volumes of unlabeled data.
• Two-step semi-supervised strategy: Trains convolutional neural networks (CNNs) on labeled data to generate predictions on unlabeled data per label type and uses those predictions as pseudo-labels to train a transformer with label-specific heads.
• Knowledge distillation: Distills task-specific CNN knowledge into a transformer model to enable simultaneous learning of all partial labels and to improve predictive accuracy and generalizability.
• Benchmarking on cardiac CT: Applied to federated cardiac CT analysis involving 8,104 cases across eight hospitals and reported performance improvements relative to UNet-based models.

Scientific Applications:

• Medical imaging diagnostics: Enables development of federated, semi-supervised models for imaging tasks such as cardiac CT analysis using mixed labeled and unlabeled datasets.
• Cross-institutional model development: Facilitates collaborative model training across hospitals to leverage diverse datasets while maintaining data locality and privacy.

Methodology:

Federated learning across institutions combined with a two-step semi-supervised method: train CNNs on labeled data, generate pseudo-labels for unlabeled data per label type, train a transformer with label-specific heads, and distill CNNs into the transformer.