FedACT: Federated Agnostic Learning on Limited Decentralized CT Images With Knowledge Transferring Process

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Informatics

Background:

• Federated learning typically trains models for specific tasks.
• Real-world clinical settings involve diverse and varying diagnostic tasks across sites.
• Existing methods struggle with heterogeneous client-side diagnostic tasks.

Purpose of the Study:

• To address the challenge of Federated Agnostic Learning (FAL) with varying client-side diagnostic tasks.
• To introduce a novel method, FedACT, for effective federated learning in agnostic task settings.
• To enhance model generalizability and performance on diverse clinical diagnostic tasks.

Main Methods:

• FedACT utilizes an end-to-end similarity layer with contrastive learning to extract shared features across agnostic tasks.
• Personalized task-specific branches (classification, segmentation) are designed for comprehensive task accomplishment.
• Specialized updating and aggregation methods are developed to handle data heterogeneity and unseen tasks.

Main Results:

• FedACT demonstrates effectiveness in various scenarios within the FAL setting.
• The method successfully extracts shared features, enhancing generalizability.
• Personalized branches achieve accurate classification and segmentation through knowledge transfer.

Conclusions:

• FedACT provides a robust solution for federated learning with agnostic client-side tasks.
• The proposed approach improves performance in diverse and heterogeneous clinical diagnostic scenarios.
• FedACT enhances model adaptability and accuracy for varied medical imaging tasks.