Federated cross-source learning for lung nodule segmentation with data characteristic-aware weight optimization

Abstract

Federated learning enables multiple medical institutions to undertake distributed training while protecting patient privacy. Nevertheless, the significant variance in data distributions across diverse sites results in imbalanced knowledge acquisition, thereby affecting the performance of the global model. To tackle this challenge, we propose a novel federated algorithm for lung nodule segmentation, incorporating a Cross-source Learning (CSL) method. This method generates pseudo nodules by synthesizing the nodule phase spectrum with the nodule amplitude spectrum from other clients. These pseudo nodules are subsequently embedded into pulmonary regions to augment the data. By incorporating knowledge from various clients, which alleviates the challenges posed by non-IID data. On the server side, a Data Characteristic-aware Weight Optimization (DCWO) method is proposed to incorporate client data quality assessment and the size of lung nodule volume as weights to optimize both model performance and fairness. On the client side, we design a Multi-scale Attention Dynamic Convolution (MADC) lightweight network, which dynamically adapts attention to different spatial regions and extracts features at multiple scales. The performance of our method is superior to the state-of-the-art methods on six public and in-house CT datasets of lung cancer.