Federated Partially Supervised Learning With Limited Decentralized Medical Images

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Data governance in healthcare necessitates privacy-preserving methods like federated learning (FL).
• Decentralization's impact on partially supervised learning, especially with limited data, is not well understood.
• Medical domain often faces data scarcity and limited partial labels per client.

Purpose of the Study:

• To formulate and discuss federated partially supervised learning (FPSL) for decentralized medical images with partial labels.
• To investigate the impact of decentralized, partially labeled data on deep learning models.
• To propose a robust framework, FedPSL, addressing FPSL challenges.

Main Methods:

• Formulated FPSL for multi-label classification using decentralized, partially labeled medical image data.
• Analyzed challenges within the FedAVG framework.
• Developed FedPSL with task-dependent model aggregation and task-agnostic decoupling learning modules.

Main Results:

• Empirical validation of FPSL as an under-explored problem with practical value.
• FedPSL demonstrated robust performance against baseline methods.
• Effectiveness shown in scenarios with data scarcity and domain shifts.

Conclusions:

• FPSL is a critical area for developing AI in decentralized medical settings with limited data.
• The proposed FedPSL framework effectively handles data scarcity and domain shifts.
• This work opens new research avenues in label-efficient learning for medical images.