Communication Efficient Federated Learning for Multi-Organ Segmentation via Knowledge Distillation With Image...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Federated learning (FL) is popular for multi-organ segmentation in CT scans.
• Current FL methods require frequent parameter exchange, posing practical challenges due to network variability and data transmission.
• Data heterogeneity, including partial labels from clients, further complicates FL efficiency.

Purpose of the Study:

• To propose an efficient communication approach for federated learning (FL) in multi-organ segmentation, specifically addressing partial labels.
• To reduce the communication overhead associated with traditional FL methods.
• To enhance the accuracy and practicality of FL for medical image segmentation.

Main Methods:

• Implemented a novel FL approach transmitting local model parameters once to a central server.
• Employed knowledge distillation (KD) to train the global model using local models.
• Generated synthetic images from client models to mitigate data distribution shifts for KD.
• Allowed for optional few-round fine-tuning using existing FL algorithms.

Main Results:

• The proposed method significantly reduces communication rounds compared to standard FL.
• Knowledge distillation with synthetic data generation effectively addresses data shifts.
• The approach demonstrates substantial performance improvements over state-of-the-art methods in few-communication scenarios.
• Evaluations on public datasets confirm the efficacy of the proposed method.

Conclusions:

• The developed efficient communication strategy enhances federated learning for multi-organ CT segmentation.
• The integration of knowledge distillation and synthetic data generation offers a practical solution for heterogeneous data and limited communication.
• This method provides a flexible and high-performing alternative for decentralized medical image analysis.