The codes for the work "ACFL: Communication-Efficient adversarial contrastive federated learning for medical image segmentation". Our paper has been accepted by Knowledge-Based Systems. We provide both command-line interface and a user-friendly web interface for easy deployment and visualization.

• 🔥 Multiple Federated Learning Methods Implementation:

  • FedAvg: Vanilla Federated Averaging
  • FedProx: Federated Optimization with Proximal Terms
  • SCAFFOLD: Stochastic Controlled Averaging
  • RSC: Representation Self-Challenging
  • ACFL: Our Proposed Adversarial Contrastive Method

• 📊 Three Dataset Support for Non-IID Scenarios:

  • PACS: A domain generalization benchmark with 7 categories across 4 domains (Photo, Art, Cartoon, Sketch)
  • OfficeHome: 65 categories across 4 diverse domains for domain adaptation
  • TerraInc: Wildlife recognition dataset with 10 categories demonstrating real-world distribution shifts

• 🌈 Web Interface: Easy-to-use GUI for:

  • Model training configuration
  • Real-time training visualization
  • TensorBoard log management

• 📈 Non-IID Feature Distribution Bias:

  • Handles domain shift challenges
  • Addresses feature distribution skew
  • Supports cross-domain generalization

• 🛠️ Easy Deployment: Simple setup and configuration

# Core Dependencies
torch==1.11.0
torchvision==0.12.0
flask==2.0.1
tensorboard==2.11.0

1. Create directories for log files
2. Update dataset paths in configs/default.py:
   • datasetpath
     • pacs_path
     • officehome_path
     • terrainc_path
   • log_count_path

You can refer to the run command in the run.sh file for startup.

# Start training/testing
sh run.sh

Or use this simple app to start and log for visualization.

# Start the web server
python app.py

Then open your browser and visit: http://localhost:5000

• 🎯 Training Configuration
  • Dataset selection
  • Model parameter adjustment
  • Training progress monitoring
• 📊 TensorBoard Integration
  • Real-time metrics visualization
  • Log management
  • Performance analysis

Sample Start

Tensorborad Visualization

Tensorborad Visualization

This project has benefited from the following excellent resources:

• FACT
• DomainBed
• FedNova
• SCAFFOLD-PyTorch

@article{liang2024acfl,
  title={ACFL: Communication-Efficient adversarial contrastive federated learning for medical image segmentation},
  author={Liang, Zhiwei and Zhao, Kui and Liang, Gang and Wu, Yifei and Guo, Jinxi},
  journal={Knowledge-Based Systems},
  volume={304},
  pages={112516},
  year={2024},
  publisher={Elsevier}
}

This project is licensed under the MIT License - see the LICENSE file for details.