Why Surgery Is Not Enough. Non-small cell lung cancer (NSCLC) accounts for approximately 85% of all lung cancers, and surgical resection with lymph node dissection is the primary treatment for early-stage disease. Despite surgery, postoperative progression -- including local recurrence and distant metastasis -- occurs in 18 to 34% of patients. Identifying who will relapse after surgery is critical for deciding whether to add adjuvant chemotherapy or intensify follow-up.
The TNM Staging Limitation. The current standard for prognostic assessment, the TNM staging system, cannot reliably predict outcomes at the individual patient level. Patients with identical TNM stages often experience dramatically different outcomes due to heterogeneity in tumor biology, imaging characteristics, and immune environment. A more granular, data-driven prognostic tool is needed to complement TNM staging for personalized treatment decisions.
The Privacy Problem in Multi-Center AI. Building accurate deep learning models for lung cancer prognosis requires large datasets from multiple hospitals. However, sharing patient CT scan data across institutions raises serious privacy and ethical concerns. This fundamental tension between data needs and privacy has limited the development of large-scale, generalizable AI models in healthcare.
Federated Learning as the Solution. Federated Learning (FL) enables multiple hospitals to collaboratively train a shared AI model without sharing raw patient data. Each center trains locally on its own data, and only model parameters (not images or patient information) are shared with a central server for aggregation. This privacy-preserving approach allows AI models to benefit from the diversity of multicenter data while respecting regulatory constraints.
Study Population and Data Sources. The study retrospectively analyzed 926 patients from four Chinese medical centers diagnosed with solid-type stage I or II NSCLC who underwent surgical resection between January 2014 and September 2019. All patients had preoperative CT scans within one month of surgery, pathological confirmation of NSCLC, and at least three years of follow-up. Patients were classified as lung cancer progression (LCP) if they experienced disease progression within three years, or non-progression (LCNP) if disease-free for the same period.
Key Innovation: Fourier-Domain Parameter Aggregation. Standard federated learning fails in medical settings because directly averaging model parameters from different hospitals can destroy each center's locally learned features. The RFed model addresses this through Exponential Progressive Fourier Aggregation (EPFA), which uses Fourier Transform to separate model parameters into low-frequency components (representing generalizable patterns shared across centers) and high-frequency components (capturing center-specific imaging characteristics). Only low-frequency parameters are aggregated globally, while high-frequency parameters remain localized.
Handling Class Imbalance. Approximately 20% of NSCLC patients experience postoperative progression -- meaning the dataset is heavily imbalanced, with far more non-progression than progression cases. Standard cross-entropy loss functions perform poorly on imbalanced data. RFed addresses this with a margin-based cross-entropy loss that calibrates decision boundaries for each class, combined with Contrastive Predictive Coding (CPC) to align local and global model representations and prevent majority-class dominance.
Three-Stage Transfer Learning. A three-stage A-distance-based Transfer Learning (ATTL) strategy smoothly transfers knowledge from the global to local models. In Stage 1, the local model learns from the global model. In Stage 2, local and global models engage in mutual learning when the global model reaches certain performance thresholds. In Stage 3, the improved global model trains back into the local model. This prevents performance regression that commonly occurs when global model updates overwrite locally specialized knowledge.
Feature Extraction and Classification. The trained federated model generates 7,616 feature maps per patient from CT images. The maximum relevance minimum redundancy (mRMR) algorithm selects the most discriminative, non-redundant features. A Bayesian extreme learning machine then performs the final binary classification (progression vs. non-progression). The base model is ResNet34, with CT lesion images standardized to 128x128 pixels after radiologist annotation reviewed by a second independent radiologist.
Performance vs. Clinical Models. The RFed model achieved AUC values of 0.936, 0.861, 0.925, and 0.970 across the four test centers, with an average AUC of 0.923. This represents a 22.7% improvement over the traditional clinical model (average AUC 0.696), which used longest tumor diameter, smoking history, lobulated sign, spiculated sign, and CEA level. DeLong tests confirmed statistically significant AUC differences (p values of 0.052, 0.012, 0.004, and 0.001 across centers), and Integrated Discrimination Improvement (IDI) values were consistently significantly positive (p less than 0.001 for all centers).
Outperforming All Other Federated Learning Methods. RFed was compared against five state-of-the-art federated learning algorithms: FedAvg, FedBn, FedProx, MetaFed, and PrrFed. RFed achieved the highest average AUC of 0.923, representing an improvement of 2.90 to 5.84% over competing methods. Competing FL methods achieved average AUCs of 0.872 (FedAvg), 0.897 (FedBn), 0.884 (FedProx), 0.884 (MetaFed), and 0.875 (PrrFed).
Clinical Utility via Decision Curve Analysis. Decision Curve Analysis (DCA) showed that RFed provided greater net clinical benefit than the traditional clinical model across threshold probabilities ranging from 0.02 to 0.99 at all four centers. This wide threshold range indicates that RFed is clinically useful across a broad range of clinical risk tolerances, not just at specific operating points.
Independent Prognostic Value. Multivariate Cox regression analysis confirmed that RFed is an independent prognostic factor for postoperative progression after adjusting for age, gender, longest diameter, and CEA. The univariate hazard ratio was 4.034 (95% CI: 2.189 to 7.435, p less than 0.001), and the multivariate hazard ratio was 9.176 (95% CI: 2.835 to 29.693, p less than 0.001). Kaplan-Meier curves showed significantly clearer separation between high-risk and low-risk groups for RFed than for clinical models at all four centers.
Out-of-Distribution Generalization. To test whether RFed truly generalizes rather than overfitting to training center characteristics, out-of-distribution experiments trained models on three centers and tested on the fourth. AUC values across all configurations ranged from 0.770 to 0.818, with F1 scores from 0.429 to 0.473. These consistent results across all train-test combinations confirm the model's ability to perform well on hospitals that contributed no training data.
Architecture Independence. The effectiveness of the RFed framework was validated across four different deep learning backbone architectures: ResNet34, VGG16, Inception-V3, and Vision Transformer (ViT). All achieved strong performance with average AUC values ranging from 0.882 (ViT) to 0.919 (ResNet34). This architecture independence demonstrates that RFed's improvements come from the federated learning framework itself rather than from the specific neural network design.
Five-Fold Cross-Validation. Cross-validation of the multicenter RFed model yielded average AUC values of 0.919, 0.858, 0.920, and 0.941 across the four test centers -- results closely matching the primary analysis and confirming the model's stability. A calibration slope of 0.892 (close to the ideal value of 1.0) confirmed that predicted probabilities are well-calibrated and reliable for clinical use.
Sensitivity at Different Specificity Thresholds. At 80% specificity, all four centers achieved greater than 85% sensitivity (Center D: 85.7%), demonstrating strong diagnostic utility for clinical settings that require high sensitivity. At the more stringent 90% specificity threshold, performance varied (Center A: 78.6% sensitivity; Centers B and C: 43.8% and 36.4%), indicating room for improvement with larger datasets. Generalizability to other cancer types was also demonstrated by testing RFed on skin cancer data, where it achieved a mean AUC of 0.908 across all centers.
t-SNE Feature Visualization. To assess the quality of learned feature representations, t-SNE dimensionality reduction was applied to both raw CT image data and RFed-processed features. Raw data showed significant overlap between the lung cancer progression (LCP) and non-progression (LCNP) classes, particularly at Centers B and C. After RFed processing, feature representations showed substantially improved class separation, demonstrating that the federated learning model learns clinically meaningful discriminative features that distinguish high-risk from low-risk patients.
Why Fourier Aggregation Outperforms Direct Averaging. Conventional federated learning aggregates model parameters by simple averaging across centers. This fails when centers have heterogeneous imaging protocols, equipment, or disease prevalence -- the dominant pattern averages out center-specific discriminative features. By separating parameters in the frequency domain and sharing only low-frequency components (which capture broad, generalizable patterns), RFed preserves center-specific high-frequency features while still benefiting from collaborative training.
Computational Resources and Practicality. The model was implemented using PyTorch 2.0 on a system with an NVIDIA RTX 4090 GPU (24GB memory) and an Intel Core i9-13900K CPU. Communication occurs every 50 training iterations, with each local model performing 5 optimization steps between communications. This communication schedule balances the need for knowledge sharing against the computational overhead of frequent server communication -- an important consideration for real-world hospital deployments with network constraints.
Clinical Impact for Treatment Decision-Making. The RFed model's primary clinical value is its ability to identify early-stage NSCLC patients at high risk of postoperative progression, enabling stratified management. High-risk patients identified by the model could receive adjuvant chemotherapy, more frequent imaging follow-up, or early enrollment in clinical trials. Low-risk patients might safely avoid the side effects and costs of aggressive adjuvant treatment. This precision approach goes beyond what standard TNM staging can offer.
Privacy-Preserving Multicenter Collaboration. The study demonstrates that federated learning enables hospitals to collaboratively build more accurate AI models than any individual center could achieve alone, without compromising patient data privacy. This is particularly important for rare disease subgroups where any single center has insufficient data for robust model training. The RFed framework is therefore not just a technical achievement but a model for how AI development in oncology can proceed ethically.
Limitations and Future Work. The study uses retrospective data, which may not fully represent the variability encountered in prospective clinical practice. The test set at each center is relatively small, limiting confidence at stringent specificity thresholds. Future work should expand the patient cohort, incorporate prospective validation, and explore integration of genomic and molecular data alongside CT imaging features. External validation at institutions outside China would further confirm global applicability.
Broader Potential. The RFed framework's successful application to skin cancer data alongside lung cancer confirms its generalizability beyond a single cancer type. The combination of privacy-preserving federated learning, Fourier-domain parameter aggregation, and class-imbalance handling could be applied to any medical imaging task involving multicenter data with heterogeneous distributions -- making it a broadly applicable framework for AI-powered clinical decision support.