[Paper Review] Experiments of Federated Learning for COVID-19 Chest X-ray Images

Boyi Liu, Bingjie Yan+3 more|arXiv (Cornell University)|Jul 5, 2020

COVID-19 diagnosis using AI4 references52 citations

TL;DR

This paper evaluates federated learning (FL) for COVID-19 chest X-ray classification using four models (COVID-Net, ResNet18, ResNeXt, MobileNet-v2) and compares FL with centralized training on the COVIDx dataset.

ABSTRACT

AI plays an important role in COVID-19 identification. Computer vision and deep learning techniques can assist in determining COVID-19 infection with Chest X-ray Images. However, for the protection and respect of the privacy of patients, the hospital's specific medical-related data did not allow leakage and sharing without permission. Collecting such training data was a major challenge. To a certain extent, this has caused a lack of sufficient data samples when performing deep learning approaches to detect COVID-19. Federated Learning is an available way to address this issue. It can effectively address the issue of data silos and get a shared model without obtaining local data. In the work, we propose the use of federated learning for COVID-19 data training and deploy experiments to verify the effectiveness. And we also compare performances of four popular models (MobileNet, ResNet18, MoblieNet, and COVID-Net) with the federated learning framework and without the framework. This work aims to inspire more researches on federated learning about COVID-19.

Motivation & Objective

• Motivate privacy-preserving collaborative learning for COVID-19 CXR classification across hospitals.
• Investigate FL performance for four networks on COVID-19 CXR data.
• Assess trade-offs between FL and non-FL in terms of convergence and accuracy.

Proposed method

• Define a federated learning setup with 5 agents and 0.4 participation per round.
• Train four models (COVID-Net, ResNet18, ResNeXt, MobileNet-v2) under FL using PyTorch.
• Use Adam optimizer with learning rate 2e-5 and weight decay 1e-7.
• Dataset: COVIDx (15,282 images) with Normal, non-COVID pneumonia, and COVID-19 labels.
• Compare FL-trained models against centralized training baselines.
• Evaluate performance (accuracy per set and per label) and provide Grad-CAM++ visual explanations.

Experimental results

Research questions

• RQ1 Can federated learning achieve comparable accuracy to centralized training on COVID-19 CXR classification?
• RQ2 Which model yields the best accuracy under FL for COVID-19 CXR images?
• RQ3 How does FL affect convergence speed and label-wise performance for COVID-19 detection?

Key findings

• ResNet18 achieves the highest accuracy on both training (96.15%) and testing (91.26%) under FL.
• COVID-Net under FL achieves 89.17% testing accuracy and 92.40% training accuracy.
• ResNeXt under FL achieves 90.37% testing accuracy and 94.66% training accuracy.
• MobileNet-v2 under FL achieves 86.83% testing accuracy and 91.16% training accuracy.
• Under centralized training, the four models show varying performance, with ResNet18 still performing strongest overall among FL and non-FL settings.
• MobileNet-v2 has the fewest parameters among the four models.