Federated Learning Algorithm Evaluation System for Heterogeneous Medical Image Data

Medical image processing requires learning meaningful features from diverse datasets, which are often private and scattered across different institutions. This leads to image data heterogeneity, including feature skew and quantity skew. Federated learning (FL) can guarantees data analysis without exposing privacy, making it ideal for healthcare. However, there is no metric available to measure the data heterogeneity, determining the appropriate FL algorithm for different scenarios is unworkable. To address this, a FL algorithms evaluation system for heterogeneous medical image is proposed, where entropy is used to quantify the data heterogeneity among clients. After normalization and segmentation of the entropy values across clients, the evaluation system is constructed for candidate algorithms, providing selection strategies incorporating the performance and efficiency. Experiments on the distributed X-ray datasets for breast cancer samples figure it out that, the proposed evaluation system can effectively screen out the most suitable FL algorithm for the current data distribution.