Federated brain tumor segmentation: An extensive benchmark.

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Privacy-preserving machine learning

Background:

• Federated learning (FL) is gaining traction in medical imaging for multi-center data aggregation while preserving privacy.
• Existing FL schemes are categorized into global, personalized, and hybrid approaches.
• The performance of these FL schemes on the Federated Brain Tumor Segmentation 2022 (FeTS2022) dataset remains unexplored.

Purpose of the Study:

• To conduct an extensive benchmark of federated learning algorithms across global, personalized, and hybrid categories for brain tumor segmentation.
• To evaluate the applicability and performance of various FL methods on the FeTS2022 dataset.
• To investigate the impact of data distribution (IID and limited data setups) on FL performance in this context.

Main Methods:

• Benchmarking of diverse federated learning algorithms (global, personalized, hybrid) on the FeTS2022 dataset.
• Evaluation of standard Federated Averaging (FedAvg) and advanced FL techniques.
• Analysis of FL performance under Independent and Identical Distributed (IID) and limited data distribution scenarios.

Main Results:

• Standard FedAvg demonstrates strong performance on the FeTS2022 task.
• Certain FL methods from each category offer marginal performance gains over FedAvg.
• Some advanced FL techniques show potential in mitigating model bias towards dominant data distributions.
• Federated learning behavior is further understood through IID and limited data distribution analyses.

Conclusions:

• Federated learning is a viable approach for multi-center brain tumor segmentation, with potential for performance enhancement beyond standard methods.
• The choice of FL strategy and data distribution significantly influences model performance and bias.
• This benchmark provides valuable insights for applying federated learning in medical imaging tasks.