FDCTGAN: Client-local CTGAN-augmented federated learning for imbalanced breast cancer classification

Abstract

Abstract Objective Federated learning enables collaborative medical model development without centralizing patient records; however, client-level class imbalance may reduce sensitivity to malignant cases. This study evaluated FDCTGAN, a client-local CTGAN-augmented federated learning framework for imbalanced breast cancer classification. Methods In FDCTGAN, each eligible client trained a local conditional generative adversarial network and generated synthetic malignant records equal to 50% of its local class deficit. Raw and synthetic data remained on the originating client. Only feature extrema were shared once per outer fold, while classifier parameters were exchanged through Federated Averaging. Five-fold stratified cross-validation on the Wisconsin Diagnostic Breast Cancer dataset compared FDCTGAN with Vanilla FL, Local SMOTE-FL, Centralized CTGAN-FL, and Centralized Pooled learning. Each augmentation method was evaluated using three augmentation seeds and three classifier seeds, producing nine stochastic runs per fold. Results Under full-batch local training, FDCTGAN increased malignant-class recall from 0.9448 ± 0.0206 with Vanilla FL to 0.9632 ± 0.0117. However, the paired recall difference was not statistically significant using a two-sided Wilcoxon signed-rank test (p = 0.125; five folds), and an increase in false positives accompanied the improvement. With a batch size of 32, FDCTGAN no longer outperformed Vanilla FL in malignant recall. Local SMOTE-FL provided the strongest overall balance among the federated approaches, whereas Centralized Pooled learning achieved the highest mean accuracy and F1 score. Property-inference analysis revealed substantial information leakage from individual client updates, and CTGAN training accounted for most of the computational runtime. Conclusion FDCTGAN represents a client-local proof of concept for addressing class imbalance in federated breast cancer classification. Its sensitivity advantage was dependent on the full-batch training protocol and was not statistically significant across the five outer folds. External validation, nested hyperparameter tuning, synthetic-data fidelity evaluation, and formal privacy-preserving mechanisms are required before clinical deployment.