Related Experiment Video
Updated: Jan 13, 2026
06:48
Automated Segmentation of Cortical Grey Matter from T1-Weighted MRI Images
Published on: January 7, 2019
9.4K
Decentralized Gossip Mutual Learning (GML) for brain tumor segmentation on multi-parametric MRI.
1Department of Radiation Oncology at Columbia University Irving Medical Center, New York, NY 100032 USA.
IEEE EMBS Special Topic Conference on Data Science and Engineering in Healthcare, Medicine and Biology
|January 9, 2026
Summary
Gossip Mutual Learning (GML) offers a decentralized approach to federated learning (FL) for medical imaging, enhancing model performance and reducing communication costs compared to traditional methods.
Area of Science:
- Artificial Intelligence
- Medical Imaging
- Distributed Systems
Background:
- Federated Learning (FL) facilitates collaborative training across medical institutions without data sharing.
- Centralized FL models face risks from server failures and performance degradation due to data heterogeneity.
- Decentralized approaches are needed to overcome limitations of traditional FL.
Purpose of the Study:
- To introduce Gossip Mutual Learning (GML), a novel decentralized FL framework.
- To enhance model performance and reduce communication overhead in medical imaging tasks.
- To address data variations across clinical sites through local model optimization.
Main Methods:
- Implemented a decentralized framework using Gossip Protocol for peer-to-peer communication.
- Incorporated mutual learning to optimize local models at each site.
- Evaluated performance on tumor segmentation using the BraTS 2021 dataset (146 cases, 4 sites).
Main Results:
- GML demonstrated superior performance compared to local models.
- GML achieved performance comparable to the FedAvg (Federated Averaging) approach.
- GML reduced communication overhead by 75% compared to traditional FL methods.
Conclusions:
- GML provides an effective decentralized alternative to traditional FL for medical applications.
- The framework addresses server failure risks and improves performance on heterogeneous data.
- GML significantly reduces communication costs, making collaborative training more efficient.
