# Federated Model Aggregation via Self-Supervised Priors for Highly Imbalanced Medical Image Classification

**Authors:** Marawan Elbatel, Hualiang Wang, Robert Martí, Huazhu Fu, Xiaomeng Li  
**arXiv:** [2307.14959](https://arxiv.org/abs/2307.14959)  
**Date:** 2023-07-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2307.14959v1/

## Key points

- Federated learning is crucial for handling highly imbalanced datasets in the medical field, such as those containing skin lesions and gastrointestinal images.
- Current federated methods often focus on optimizing a global model without considering intra-class variations in medical imaging due to diverse populations, findings, and scanning devices.
- Researchers conducted a study using self-supervised auxiliary networks to explore inter-client intra-class variations and found that utilizing a shared auxiliary pre-trained model locally on each client resulted in consistent divergence measurements.
- A dynamic balanced model aggregation technique called Self-Supervised Priors (MAS) was developed to guide the optimization of the global model, offering flexibility to be combined with various local learning methods for effective model aggregation.
- The study by Marawan Elbatel, Hualiang Wang, Robert Martí, Huazhu Fu, and Xiaomeng Li emphasizes the importance of considering intra-class variations in federated learning within the medical imaging domain.
- The code for their research is openly accessible at https://github.com/xmed-lab/Fed-MAS.

## Summary

In the medical field, federated learning is a crucial approach for handling highly imbalanced datasets such as those containing skin lesions and gastrointestinal images. Current federated methods often focus on optimizing a global model without considering the intra-class variations that can occur in medical imaging due to diverse populations, findings, and scanning devices. To address this issue, researchers conducted a study using publicly available self-supervised auxiliary networks to explore inter-client intra-class variations. The study revealed that utilizing a shared auxiliary pre-trained model like MoCo-V2 locally on each client resulted in consistent divergence measurements. Building upon these findings, they developed a dynamic balanced model aggregation technique called Self-Supervised Priors (MAS) to guide the optimization of the global model. Fed-MAS offers flexibility to be combined with various local learning methods to achieve effective model aggregation towards creating a highly robust and unbiased global model. The authors of this study include Marawan Elbatel, Hualiang Wang, Robert Martí, Huazhu Fu, and Xiaomeng Li. Their work titled "Federated Model Aggregation via Self-Supervised Priors for Highly Imbalanced Medical Image Classification" sheds light on the importance of considering intra-class variations in federated learning within the medical imaging domain. The code for their research is openly accessible at https://github.com/xmed-lab/Fed-MAS. This refined summary emphasizes how leveraging self-supervised auxiliary networks and dynamic balanced model aggregation techniques can enhance the performance and reliability of global models in highly imbalanced medical image classification tasks.