FedIOD: Federated Multi-Organ Segmentation From Partial Labels by Exploring Inter-Organ Dependency.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Multi-organ segmentation is crucial but typically requires fully labeled data.
• Data privacy concerns and incomplete labels hinder current segmentation approaches.
• Federated learning (FL) addresses privacy, but FL with partial labels remains underexplored.

Purpose of the Study:

• To develop a federated learning framework for multi-organ segmentation that effectively handles partial labels.
• To propose a method for generating pseudo full supervision by leveraging inter-organ dependencies from partial labels.
• To enhance feature extraction in a federated setting, mitigating bias towards specific organs.

Main Methods:

• Introduced FedIOD, a single-encoder-multi-decoder framework for federated multi-organ segmentation.
• Utilized a transformer module to model pairwise inter-organ dependency and generate pseudo full labels from partial labels.
• Employed pseudo full labels for regularization to train a shared encoder for comprehensive feature extraction.
• Each decoder was trained on its corresponding partial labels to project shared features into specific organ spaces.

Main Results:

• FedIOD demonstrated effectiveness across five diverse datasets (LiTS, KiTS, MSD, BCTV, ACDC).
• Outperformed state-of-the-art methods in in-federation evaluation for multi-organ segmentation with partial labels.
• Achieved the second-best performance in out-of-federation evaluation, showcasing robustness.

Conclusions:

• FedIOD successfully addresses the challenge of multi-organ segmentation with partial labels in a federated learning context.
• The proposed inter-organ dependency modeling effectively generates pseudo full supervision, enhancing segmentation performance.
• This framework offers a promising solution for privacy-preserving medical image segmentation when complete annotations are unavailable.