Non-IID Medical Image Segmentation Based on Cascaded Diffusion Model for Diverse Multi-Center Scenarios.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Federated learning (FL) faces challenges in multi-center medical datasets due to data heterogeneity and privacy concerns.
• Existing FL methods struggle with Non-Independent and Identically Distributed (Non-IID) data and incur high communication costs.

Purpose of the Study:

• To propose a practical privacy-preserving framework for training Non-IID medical image segmentation models in multi-center settings with low communication overhead.
• To mitigate data heterogeneity and improve the efficiency of federated learning in healthcare.

Main Methods:

• Developed an efficient cascaded diffusion model to generate synthetic image-mask pairs, addressing data heterogeneity.
• Implemented a label construction module to enhance the quality of generated data.
• Proposed aggregation methods (CD-Syn, CD-Ens, CD-KD) for diverse scenarios, balancing efficiency, privacy, and accuracy.

Main Results:

• The framework achieved an average improvement of 5.38% in Dice score compared to the baseline FednnU-Net across five Non-IID medical datasets.
• Demonstrated effectiveness in practical multi-center settings with varying demands for privacy and accuracy.
• Successfully reduced communication costs while maintaining high performance.

Conclusions:

• The proposed privacy-preserving framework offers a flexible and effective solution for Non-IID medical image segmentation using federated learning.
• The cascaded diffusion model and novel aggregation strategies significantly improve model performance and data utility.
• This approach provides a practical pathway for leveraging multi-center medical data securely and efficiently.