FedPneu: Federated Learning for Pneumonia Detection across Multiclient Cross-Silo Healthcare Datasets

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer-Aided Diagnosis

Background:

• Pneumonia is a leading global cause of mortality.
• Traditional deep learning for medical imaging raises data privacy concerns.
• Federated learning offers a decentralized approach to mitigate privacy risks.

Purpose of the Study:

• To develop and evaluate a federated deep learning model for pneumonia detection.
• To assess the performance of the FedPneu model across different client architectures.
• To enhance patient data privacy in pneumonia prediction using medical imaging.

Main Methods:

• Implementation of the FedPneu model using a federated learning framework.
• Application of the model for early pneumonia detection on X-ray images.
• Evaluation of model performance with 2, 3, 4, and 5 clients, configuring parameters like learning rate and epochs.

Main Results:

• The 2-client FedPneu architecture achieved the highest accuracy of 85.632%.
• Accuracies for 3, 4, and 5 clients were 85.536%, 76.112%, and 74.123%, respectively.
• The 2-client setup demonstrated superior performance in pneumonia detection.

Conclusions:

• The 2-client FedPneu architecture is optimal for privacy-protected pneumonia detection.
• Federated learning with multi-silo datasets enhances patient data privacy.
• The FedPneu model offers improved prediction outcomes while safeguarding sensitive health information.