HistoFL

HistoFL implements federated learning and weakly-supervised attention multiple instance learning to train deep learning models on gigapixel whole slide images (WSIs) for privacy-preserving computational pathology tasks including survival prediction and patient stratification.

Key Features:

• Federated Learning Framework: Enables collaborative model training across multiple institutions without directly sharing sensitive patient data.
• Weakly-Supervised Attention Multiple Instance Learning: Trains on slide-level labels using attention-based multiple instance learning to avoid requiring detailed cellular or tissue annotations.
• Differential Privacy: Applies differential privacy by adding randomized noise to model updates during training to protect individual patient data.
• Survival Prediction and Patient Stratification: Provides a weakly-supervised learning framework for survival prediction and patient stratification from WSIs.
• Scalability and Performance: Scales to large multi-institutional datasets, including handling datasets comprising hundreds of thousands of gigapixel images.

Scientific Applications:

• Morphological Phenotype Characterization: Identifies and characterizes known morphological features within histology images.
• Prediction of Molecular Alterations: Predicts non-human-identifiable molecular changes from histological data.
• Diagnostic Problem Solving: Evaluated on diagnostic challenges using thousands of WSIs to develop and validate models from distributed data sources.

Methodology:

Performs multi-centric data integration without direct sharing, trains federated models using weakly-supervised attention multiple instance learning with differential privacy (adding randomized noise to model updates), and evaluates models on diagnostic tasks using WSIs.