Federated Learning for Medical Image Classification: A Comprehensive Benchmark.

Area of Science:

• Medical Imaging
• Machine Learning
• Artificial Intelligence

Background:

• Federated learning (FL) is suitable for multi-center medical image analysis, preserving data privacy.
• Existing FL research often lacks comprehensive medical imaging evaluations, focusing on natural images.

Purpose of the Study:

• To comprehensively evaluate state-of-the-art FL algorithms for medical image classification.
• To assess system performance metrics like communication cost and computational efficiency in medical FL.
• To propose an improved FL method for medical imaging tasks.

Main Methods:

• Conducted a fair comparison of classification models using various FL algorithms across multiple medical imaging datasets.
• Evaluated system performance metrics including communication cost and computational efficiency.
• Developed a novel method combining generative denoising diffusion probabilistic models and label smoothing for data augmentation.

Main Results:

• Medical imaging datasets present significant challenges for current FL optimization algorithms.
• No single FL algorithm consistently achieved optimal performance across all tested medical scenarios.
• The proposed method significantly enhanced FL performance on classification tasks across various medical imaging datasets.

Conclusions:

• Current FL optimization algorithms may underperform on medical imaging datasets.
• A benchmark and guidance are provided for future FL research in medical imaging.
• The developed generative data augmentation technique offers an efficient and robust solution for improving FL in medical contexts.