FedRS: Federated Learning Under Reliable Supervision for Multi-Organ Segmentation With Inconsistent Labels

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Multi-organ segmentation typically requires large, fully labeled datasets.
• Medical image data is often decentralized due to privacy concerns and partially labeled due to high annotation costs, leading to label inconsistency.
• Federated learning enables privacy-preserving decentralized training but struggles with label inconsistency, causing model divergence.

Purpose of the Study:

• To propose an effective and communication-efficient federated learning method for multi-organ segmentation.
• To address the challenge of label inconsistency in decentralized medical image datasets.
• To improve the accuracy and generalization capability of federated learning models in medical image segmentation.

Main Methods:

• Developed Federated Learning under Reliable Supervision (FedRS) with Less-Forgetting and Less-Constraint loss functions to reduce local model divergence.
• Implemented a global model aggregation strategy based on prediction consistency between local and global models.
• Utilized a lightweight model backbone (4.1M parameters) to minimize communication costs.

Main Results:

• FedRS demonstrated superior performance compared to localized, centralized, and state-of-the-art federated learning methods on nine public 3D abdominal CT datasets.
• The method achieved strong effectiveness and generalization capabilities on both in-federation and out-of-federation datasets.
• FedRS significantly reduced communication costs due to its efficient model architecture.

Conclusions:

• FedRS effectively handles label inconsistency in federated medical image segmentation.
• The proposed method enhances global model reliability and achieves state-of-the-art performance.
• FedRS offers a communication-efficient solution for decentralized medical image analysis.