Enhancing breast magnetic resonance imaging segmentation with a federated semi-supervised approach

Area of Science:

• Artificial Intelligence in Medical Imaging
• Machine Learning for Healthcare
• Radiology and Diagnostic Imaging

Background:

• Deep learning significantly improves automatic segmentation of breast Magnetic Resonance Imaging (MRI), aiding diagnosis.
• Training deep learning models demands extensive annotated data, a challenge for individual institutions, particularly smaller ones.
• Data annotation is resource-intensive, limiting the development of robust AI models in medical imaging.

Purpose of the Study:

• To develop a federated semi-supervised learning framework for automated breast MRI segmentation.
• To maximize resource utilization and preserve data privacy across collaborating institutions.
• To enhance model robustness and generalization using limited annotated data.

Main Methods:

• Implemented a federated learning approach where each institution trains models locally.
• Utilized semi-supervised learning with perturbations on unannotated samples and features for local training.
• Designed a joint optimization loss function for both annotated and unannotated data.

Main Results:

• Achieved superior performance in breast MRI segmentation across three hospitals.
• Obtained a Dice Similarity Coefficient (DSC) of 94.8% and Intersection over Union (IoU) of 86.6%.
• Outperformed existing models by up to 8.8% (DSC) and 12% (IoU).

Conclusions:

• The proposed federated semi-supervised learning framework effectively addresses data limitations in breast MRI segmentation.
• The method enhances model performance and generalization while maintaining data privacy.
• This approach offers a scalable solution for developing advanced AI diagnostic tools in medical institutions.