A decentralized federated learning approach for predicting chemoradiotherapy response from multicenter whole-slide images

Given the institutional and regulatory barriers to centralized sharing of multicenter whole-slide image (WSI) data for non-small cell lung cancer (NSCLC), federated learning (FL) enables collaborative model development without requiring centralized sharing raw WSI data. This study explores a WSI–based FL model for predicting chemoradiotherapy response in advanced NSCLC, and compares its performance with centralized deep learning (DL) and clinical models using external multi-center validation cohorts. This study collected WSI and corresponding clinical data from 250 patients across four hospitals. We developed a centralized DL model and a FL model based on WSI, with patch-level predictions aggregated to patient-level outcomes using a voting method. In addition, we built a clinical logistic regression model using clinical data. All models were evaluated on two independent external cohorts (Hospital C and Hospital D). Performance was summarized as the mean and 95% confidence interval across the five models obtained from five-fold cross-validation. Across both external validation cohorts, WSI–based models consistently outperformed the clinical model. The FL model achieved performance comparable to the DL model. In hospital C, the DL model had an accuracy/AUC of 0.68/0.64, while the FL model achieved 0.70/0.65. The clinical model had an accuracy/AUC of 0.58/0.58. In hospital D, the DL model had an accuracy/AUC of 0.68/0.68, while the FL model achieved 0.73/0.71. The clinical model had an accuracy/AUC of 0.61/0.59. The FL model using WSI achieved performance comparable to the centralized DL model in predicting NSCLC outcomes, while allowing each participating institution to retain its local data. It also showed consistent performance across independent external validation cohorts. This approach provides a practical solution for collaborative model development in multi-center pathology studies, particularly in settings where data sharing is restricted.