Performance of federated versus centralized learning for mammography classification across film-digital domain shift.

Introduction: Large, diverse datasets are essential for reliable deep learning in mammography, yet clinical data remain siloed due to privacy and governance constraints. Federated learning enables collaborative training without sharing raw data, but its robustness under strong imaging-domain heterogeneity, such as film-digital shifts, remains uncertain. Methods: We conducted a comparative evaluation of centralized learning and cross-silo federated learning for benign-malignant lesion classification across two heterogeneous public datasets: CBIS-DDSM (scanned film) and VinDr-Mammo (full-field digital). Using ResNet-50 and Swin V2-T backbones, we evaluated FedAvg, FedProx, SCAFFOLD, and FedBN across multi-seed experiments with bootstrap confidence intervals. The study design included local-only baselines, homogeneous FL controls, size-balancing ablations, and a resolution ablation (224→324 px). Performance was assessed using AUROC, AP, Accuracy, Precision, Recall, F1, and Precision@Recall = 0.90. Results: Federated models matched centralized learning in homogeneous settings for both domains. Under film-digital heterogeneity, FL retained strong performance on the digital VinDr domain (AUC ≈ 0.91-0.95) but showed reduced performance on the film-based CBIS domain (AUC ≈ 0.53-0.62), exhibiting a shift toward high-recall/low-precision behavior. Neither FedProx, SCAFFOLD, nor FedBN consistently mitigated this degradation. Size-balancing improved CBIS performance modestly but did not close the gap to centralized learning, indicating that feature and quality shift dominated over dataset-size imbalance. Higher input resolution improved CBIS calcification detection (e.g., F1 0.49 → 0.54). Discussion: These findings show that FL performs reliably within homogeneous domains but remains vulnerable to strong feature and quality shifts between film and digital mammography. The observed asymmetric performance suggests that domain shift, rather than data quantity or optimizer instability, is the dominant limiting factor. Conclusion: Federated learning enables high-performing mammography classification without data centralization in homogeneous settings but requires domain-aware or personalized FL strategies, site-specific thresholding, and resolution-sensitive preprocessing to ensure reliable deployment under film-digital heterogeneity.