Related Experiment Video
Updated: Jan 7, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
Decentralized Gossip Mutual Learning (GML) for automatic head and neck tumor segmentation
1Department of Radiation Oncology, Columbia University Irving Medical Center 622 W 168th St, New York, NY 10032, United States.
Proceedings of Spie--The International Society for Optical Engineering
|December 30, 2025
Summary
Gossip Mutual Learning (GML) enhances decentralized medical AI by enabling peer-to-peer learning without a central server. This approach improves tumor segmentation accuracy and significantly reduces communication overhead compared to traditional federated learning methods.
Area of Science:
- Medical Imaging
- Machine Learning
- Artificial Intelligence
Background:
- Federated learning (FL) enables collaborative model training across medical centers without data sharing.
- Traditional FL's reliance on a central server presents a single point of failure and may not optimize for local data variations.
- Site-specific data heterogeneity can limit the performance of globally trained models.
Purpose of the Study:
- To introduce Gossip Mutual Learning (GML), a decentralized framework for collaborative machine learning in healthcare.
- To address the limitations of centralized FL, including server vulnerability and suboptimal local performance.
- To improve tumor segmentation accuracy on PET/CT images through a novel decentralized approach.
Main Methods:
- Developed Gossip Mutual Learning (GML), a decentralized collaborative learning framework.
- Utilized Gossip Protocol for direct peer-to-peer communication among clinical sites.
- Implemented mutual learning where each site optimizes its local model using information from peers.
- Evaluated GML on the HECKTOR21 dataset for tumor segmentation on PET/CT images.
Main Results:
- GML improved tumor segmentation performance (Dice Similarity Coefficient) by 3.2%, 4.6%, and 10.4% compared to pooled training, FedAvg, and individual training, respectively.
- GML demonstrated comparable generalization performance to pooled training and FedAvg on out-of-sample sites.
- GML achieved a sixfold decrease in communication overhead compared to FedAvg, using only 16.67% of the communication cost.
Conclusions:
- GML offers a robust and efficient decentralized alternative to traditional federated learning for medical image analysis.
- The GML framework enhances local model performance and reduces communication overhead, making it suitable for distributed medical AI applications.
- Decentralized peer-to-peer learning effectively addresses data heterogeneity challenges in multi-center medical studies.
