A privacy-preserving multicenter framework for lung nodule segmentation in low-dose CT using an F1-score-weighted federated 3D Transformer-UNet
Cheng-Shun Lee, Kai Chain, Hsueh-Hsing Pan, Li-Chun HuangBackground
Early detection of lung nodules using low-dose computed tomography (LDCT) is essential for reducing lung cancer mortality. Although deep-learning-based segmentation models have demonstrated promising performances, centralized training across multiple medical institutions is often infeasible because of patient privacy regulations and data-sharing constraints. Heterogeneous data distribution among hospitals may also degrade the robustness of collaboratively trained models. A privacy-preserving and robust multicenter segmentation framework is required for real-world clinical deployment.
Methods
We developed a privacy-preserving federated framework based on a 3D Transformer-UNet. Multi-head self-attention was applied at the network bottleneck to capture long-range volumetric dependencies, whereas multi-head cross-attention was incorporated into the skip connections to regulate the fusion of encoder and decoder features. Three physically independent client workstations participated in every communication round. Each local model was evaluated using a server-side LUNA16 reference set, and the resulting nodule-level F1 scores were normalized to determine the aggregation weights. The initially aggregated model was subsequently refined using LUNA16-assisted morphological correction.
Results
The complete framework achieved nodule-level precision, recall, and F1 scores of 84.5%, 85.1%, and 84.8%, respectively. At the voxel level, it achieved a Dice similarity coefficient of 78.9%, an Intersection-over-Union of 65.2%, and an average surface distance of 1.36 mm. The proposed method reached the predefined convergence criterion after 45 communication rounds, compared with 85 rounds for FedAvg, corresponding to a 47.1% reduction in the estimated cumulative model-transfer payload required to reach convergence.
Conclusions
The proposed framework combined attention-based volumetric feature learning, F1-score-weighted aggregation, and server-side morphological correction. Under the evaluated three-client simulation, this combination improved nodule-level detection, voxel-level segmentation, and convergence efficiency while avoiding the exchange of private LDCT images. Further validation using prospective and naturally distributed multicenter clinical data is required before clinical deployment.