FedDBL: Communication and Data Efficient Federated Deep-Broad Learning for Histopathological Tissue Classification.

Area of Science:

• Computational Pathology
• Artificial Intelligence
• Medical Imaging

Background:

• Histopathological tissue classification is crucial in computational pathology.
• Deep learning (DL) models excel but centralized training risks data privacy.
• Existing federated learning (FL) frameworks demand extensive data and communication rounds.

Purpose of the Study:

• To introduce a lightweight and universal FL framework, Federated Deep-Broad Learning (FedDBL).
• To achieve superior classification performance with minimal annotated samples and single-round communication.
• To address privacy concerns and reduce communication overhead in FL for pathology.

Main Methods:

• Integrating a pretrained DL feature extractor with a broad learning inference system.
• Utilizing a classical federated aggregation approach within the FedDBL framework.
• Implementing FedDBL with a ResNet-50 backbone for experiments.

Main Results:

• FedDBL significantly outperforms competitors with limited data and one-round communication.
• Achieves performance comparable to multi-round FL methods.
• Reduces client communication burden from 4.6 GB to 138.4 KB per client.
• Demonstrates scalability in generalization, personalization, and across image modalities.

Conclusions:

• FedDBL provides an efficient, privacy-preserving solution for histopathological classification.
• The framework drastically reduces data and communication requirements.
• Ensures model security against inversion attacks due to no data sharing.