ResNetFed: Federated Deep Learning Architecture for Privacy-Preserving Pneumonia Detection from COVID-19 Chest...

Area of Science:

• Artificial Intelligence in Healthcare
• Medical Imaging Analysis
• Privacy-Preserving Machine Learning

Background:

• Centralized machine learning in healthcare faces challenges due to personal health data privacy regulations.
• Federated Learning (FL) offers a decentralized approach, training models on siloed data to maintain privacy.
• Accurate and private methods are needed for COVID-19 detection using medical imaging.

Purpose of the Study:

• To investigate the viability of Federated Learning for detecting COVID-19 pneumonia from chest radiographs.
• To propose and evaluate a privacy-preserving federated model, ResNetFed, incorporating Differential Privacy.
• To assess the performance of the federated approach under uneven data distribution scenarios.

Main Methods:

• Utilized 1411 chest radiographs (753 normal, 658 COVID-19) from the COVIDx8 dataset.
• Partitioned data unevenly across five simulated data silos to mimic real-world FL conditions.
• Developed ResNetFed, a modified pre-trained ResNet50 model with Differential Privacy, for federated training.

Main Results:

• ResNetFed achieved a mean accuracy of 82.82%, significantly outperforming locally trained ResNet50 models (63% accuracy).
• The federated approach demonstrated superior performance, especially in underpopulated data silos, with accuracy gains up to +34.9%.
• ResNetFed effectively addressed performance disparities caused by uneven data distribution among silos.

Conclusions:

• Federated Learning, particularly with ResNetFed and Differential Privacy, is a viable and effective method for COVID-19 pneumonia detection.
• The proposed federated solution enhances privacy while maintaining high accuracy, even with imbalanced datasets.
• ResNetFed offers a practical, privacy-preserving tool to aid in initial COVID-19 screening in medical settings.