Source-linked AI summary
Experiments of Federated Learning for COVID-19 Chest X-ray Images
Boyi Liu, Bingjie Yan, Yize Zhou, Yifan Yang, Yixian Zhang
TL;DR
COVID-19 CXR detection lacks sufficient shared training data because hospitals cannot freely disclose patient medical information. This paper applies federated learning and compares four models with and without the framework, finding ResNet18 strongest overall under the reported experiments.
Problem
COVID-19 CXR detection lacks sufficient shared training data because hospitals cannot freely disclose patient medical information.
Method
Federated learning trains a shared model across institutions by aggregating local model updates without sharing local raw data.
Results
ResNet18 achieved the highest training and testing accuracy after 100 rounds, reaching 96.15% and 91.26%, respectively.
Takeaways & Limitations
The experiments indicate that ResNeXt and ResNet18 are preferable among the four evaluated models for COVID-19 identification.
Abstract
from arXiv · showhide
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.
I. INTRODUCTION
COVID-19 screening with chest X-rays can benefit from deep learning, but privacy restrictions and distributed hospital data limit access to sufficient training samples. The paper therefore evaluates federated learning for COVID-19 CXR training across four neural networks.
- Privacy restrictions prevent hospitals from sharing medical data, creating a shortage of training samples for COVID-19 detection.
- Federated learning addresses data silos by training a shared model without obtaining local patient data.
- The study conducts four federated-learning experiments on COVID-19 CXR images using CovidNet, ResNeXt, MobileNet-v2, and ResNet18.
- The experiments compare model performance under federated learning and analyze results to motivate future COVID-19 federated-learning research.
II. RELATED WORK
Related work positions federated learning as a privacy-preserving approach for distributed medical and other sensitive data. Prior COVID-19 imaging research provides the diagnostic context for applying it to CXR analysis.
- COVID-Net uses a PEPX compression structure to identify COVID-19 pneumonia in CXR images while retaining network performance and sensitivity to pneumonia features.
- Federated learning trains across devices by aggregating updated information at a central server rather than sending all training data there.
- Federated learning has been studied in financial security, artificial intelligence, and robotics before its application to COVID-19 imaging.
- Prior work also combines federated learning with meta-learning and applies it to human activity identification.
III. FEDERATED LEARNING SYSTEM FOR COVID-19 CXR IMAGES
The proposed system models multiple hospital data owners that jointly train a COVID-19 CXR classifier without sharing patient images. Federated performance is intended to approach centralized training while preserving local data separation.
- The system defines N COVID-19 CXR image owners as F1, F2, ..., FN from different hospitals.
- Because patient medical data, including CXR images, cannot be shared, each owner trains using its own local dataset.
- Federated learning jointly trains a model MFED while keeping each owner’s local dataset Di undisclosed.
- The framework considers federated learning successful when federated performance VFED is within ε of centralized performance VSUM.
B. Framework of the Federated Learning System
The federated-learning framework maintains a global model on a server while participants train local models on private data. Training proceeds through initialization, local updating, and server aggregation.
- Framework of the Federated Learning System: The server maintains the overall model, distributes it to user terminals, and updates it from uploaded local model parameters.
- Framework of the Federated Learning System: The framework assumes participating data owners are honest and provide truthful private training data.
- Framework of the Federated Learning System: The local model is trained on each participating device, whereas the global model is produced after server aggregation.
- Framework of the Federated Learning System: Initialization specifies the task, global model, and parameters before the server allocates the initialized model to clients.
- Framework of the Federated Learning System: During local training, each participant updates model parameters using local data to minimize its loss function.
- Framework of the Federated Learning System: The server aggregates participating local models and sends the updated global parameters back to data-holding users.
IV. EXPERIMENTS
The experiments evaluate COVID-19 pneumonia chest X-ray recognition using multiple models and a federated learning framework on the COVIDx dataset.
- The experiments recognize COVID-19 pneumonia in chest X-ray images using various models and federated learning frameworks.
- The COVIDx dataset contains 15,282 images, including 13,703 training images and 1,579 test images.
- COVIDx combines images from five publicly available or open-access chest X-ray datasets.
1) Model:
The study uses four neural-network models for COVID-19 chest X-ray classification, including residual, lightweight, and COVID-19-specific architectures.
- Four models are used for the image-classification experiments.
- COVID-Net uses PEPX to compress its network structure while largely preserving performance and maintaining high sensitivity to COVID-19 pneumonia characteristics.
- ResNet18 adds an identity-mapping layer to deepen the network and help prevent accuracy loss from overfitting.
- ResNeXt applies a split-transform-merge strategy that converts single-core convolution into multi-core convolution while retaining ResNet18’s topology.
- MobileNet-v2 is a lightweight convolutional neural network with a feature-processing sequence that differs from residual networks.
2) Implementation and Training:
Federated learning is implemented as pseudo-distributed training on one machine, with separate agent models synchronized to an updated central model.
- Federated learning is implemented as pseudo-distributed training completed on one machine.
- Each agent has a separate model that is reset to the updated central model after each central-model update.
- The models are implemented in PyTorch, and training and test images are resized to (224,224).
- Each terminal agent uses Adam with learning rate = 2e-5 and weight decay = 1e-7.
C. Experimental Results and Analysis
Across the experiments, the models converged under federated learning, with ResNet18 showing the fastest convergence and highest overall accuracy. Comparisons also examined label sensitivity, federated versus individual training, and parameter efficiency.
- ResNet18 converged fastest and achieved 96.15% training accuracy and 91.26% testing accuracy after 100 rounds.
- ResNeXt followed ResNet18 in convergence speed, while COVID-Net ranked second in accuracy.
- MobileNet-v2 had a loss value similar to COVID-Net but unsatisfactory testing-set accuracy.
- The experiments evaluated four models’ performance and label sensitivity using federated learning, including convergence, accuracy, and per-label results.
- Federated learning slightly reduced the loss convergence rate compared with training without the framework.
- MobileNet-V2 showed a larger federated-versus-individual accuracy gap than ResNet18.
- MobileNet had the fewest parameters and lowest accuracy, whereas ResNet18 had the second-fewest parameters and highest accuracy.
- Grad-CAM++ visual explanations compared COVIDNet, ResNet18, ResNeXt, and MobileNet-v2 under federated and non-federated training.
V. CONCLUSION
The experiments compare four models for COVID-19 identification with chest X-ray images under federated and non-federated training. ResNet18 and ResNeXt show the strongest overall suitability, while MobileNet has the fewest parameters.
- ResNet18 has the best performance with and without the federated learning framework.
- ResNeXt performs best on images with COVID-19 labels.
- MobileNet has the fewest parameters among the evaluated models.
- ResNeXt and ResNet18 are indicated as better choices for COVID-19 identification among the four models.