Low Heterogeneity Feature Transfer-Based Federated Learning Algorithm for Early Stage Lung Cancer Classification

The problem of imbalanced distribution in multi-center medical data leads to suboptimal model performance. Therefore, to address the heterogeneity issue in multi-center data, a low heterogeneity feature transfer-based federated learning algorithm is proposed. In this algorithm, considering the influence of heterogeneous feature information in the global model, a low heterogeneity feature selection network (LHFS-Net) is introduced to construct local models and reduce the impact of heterogeneity features on local model performance. Additionally, an dynamic aggregation network (DA-Net) is proposed to filter heterogeneous feature information from each local model and aggregate the global model, thereby improving its performance. Moreover, to validate the performance of the proposed algorithm, the study collected early-stage lung cancer data from four medical institutions for the classification of early stage lung cancer medical images. The experimental results demonstrate that the proposed algorithm outperforms three existing methods across the four early stage lung cancer data centers, with AUC values of 0.8108, 0.7894, 0.7467, and 0.7527 for each respective test set.