SFL-DCSA: Split Federated Learning for Breast Cancer Prediction with Dynamic Client Selection Aggregation

Accompanied by the booming development of artificial intelligence technology, deep learning has been widely used in many fields of cancer, specially in cancer prediction. The dependence on training data for deep learning naturally raises privacy leakage concerns. Federated learning is a promising solution to these issues, but it is limited by the computational capacity of clients. Therefore, this paper has proposed a Split Federated Learning (SFL)-based breast cancer prediction scheme with dynamic client selection aggregation, by aggregating data from multiple healthcare organizations under the premise of privacy protection. First, the split learning is integrated into federated learning to predict breast canner, which contributes to protecting data privacy and reducing the computational burden on client devices. Then, the dynamic client selection aggregation is devised to lower the aggregation communication costs and improve communication efficiency by utilizing Long Short-Term Memory (LSTM) networks to evaluate the availability of each client devices in participating training. Finally, we have conducted extensive experiments on the CAMELYON16 dataset to evaluate the performance of our proposed scheme, and the experimental results have shown that our proposed scheme can converge faster and achieve the lower communication costs.