Federated Online Learning With Filters for Lung Nodule Segmentation in Low-Dose Computed Tomography Images

Lung nodules are critical indicators of pulmonary lesions closely associated with lung cancer progression. Low-dose computed tomography (LDCT) is widely employed to identify lung nodules. However, the large volume of imaging data it produces creates significant processing challenges for healthcare professionals, leading to diagnostic delays, uncertainty, and increased labor costs. To address these challenges, we propose a federated online learning framework that enables collaboration among multiple hospitals while complying with regulations. Our system includes three filters for lung nodule segmentation: a relabeling data balance filter, a two-stage network filter for segmentation and false positive reduction, and an interactive labeling correction filter. The system improves model robustness and mitigates biases during collaborative training. Tested in three hospitals, our framework preserved data privacy and achieved an accuracy of approximately 85%, outperforming models trained on local data.