Radiomic Evaluation Model on the Efficacy of Neoadjuvant Chemotherapy for Non-small Cell Lung Cancer A Multicenter Collaborative Research Based on Privacy Protection

Background: Practical implementation of radiomics research faces significant data accessibility challenges due to privacy and ethical restrictions on multicenter data aggregation. Federated Learning (FL) provides a secure distributed framework that preserves data privacy through cryptographic techniques. Its adoption in radiomics is an emerging trend, enabling collaborative training without sharing sensitive imaging data. However, the inherently Non-IID data distribution across clients in FL often leads to class imbalance, which can substantially degrade global model performance. Purpose: To develop a privacy-preserving, multicenter collaborative CT-radiomics model for evaluating neoadjuvant chemotherapy efficacy in non‑small cell lung cancer (NSCLC). Methods: To mitigate FL performance degradation caused by data imbalance, we propose a parameter‑sharing federated aggregation algorithm (FedPS), where model parameters are sequentially shared via the server. Results: On an imbalanced NSCLC NAC efficacy dataset, centralized learning achieved an AUC of 0.92. FedPS attained competitive performance (AUC = 0.88), approaching the centralized benchmark while preserving privacy. Common FL algorithms performed lower: FedAvg (AUC = 0.84), FedSGD (0.85), and FedProx (0.85). On extremely imbalanced data, FedPS maintained good performance (AUC = 0.86), compared to FedAvg (0.80), FedSGD (0.83), and FedProx (0.85). Conclusions: The proposed FedPS algorithm demonstrates promising classification and generalization performance in imbalanced federated learning scenarios.