Proximal guided hybrid federated learning approach with parameter efficient adaptive intelligence for pneumonia...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Federated Learning for Healthcare
• Computer-Aided Diagnosis

Background:

• Pneumonia diagnosis relies on chest radiography, but data privacy and fragmentation hinder AI development.
• Federated Learning (FL) enables collaborative AI training without sharing sensitive patient data.
• Existing FL methods like FedAvg struggle with data heterogeneity and high communication overhead.

Purpose of the Study:

• To propose an enhanced federated learning framework for pneumonia detection in resource-constrained settings.
• To address challenges of data heterogeneity and communication efficiency in federated medical image analysis.
• To improve the accuracy and practicality of AI-driven pneumonia diagnosis.

Main Methods:

• Implemented an upgraded federated framework using FedProx for proximal optimization and Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning.
• Utilized Vision Transformers (ViT) as the backbone for chest X-ray classification, capturing global context.
• Simulated a distributed environment with the Chest X-Ray Images dataset across multiple clients.

Main Results:

• Achieved 88.5% classification accuracy for pneumonia detection under data heterogeneity.
• Significantly reduced communication overhead and computational costs compared to standard FL methods.
• Attention heatmaps demonstrated clinical relevance, highlighting important pulmonary areas.

Conclusions:

• The proposed federated learning approach enhances pneumonia detection accuracy and efficiency.
• The framework is suitable for resource-constrained medical infrastructure due to its low memory footprint.
• Explainability features increase trust and transparency, facilitating clinical adoption.