Abstract
Breast cancer is a critical global health challenge, and deep learning shows transformative potential for medical image classification. However, privacy regulations such as HIPAA and GDPR create barriers to centralized data aggregation across institutions. This paper presents an empirical evaluation of federated learning (FL) for breast cancer classification in ultrasound images, systematically comparing seven deep learning architectures (ResNet-50, VGG16, VGG19, DenseNet-121, MobileNetV2, Vision Transformer, CoAtNet) across three FL algorithms (FedAvg, FedProx, FedOpt) with client-side differential privacy (DP). Using a simulated federation of eight institutions, we evaluate three clinically relevant classification scenarios. Federated models achieve performance comparable to centralized baselines-98.52% accuracy for normal/abnormal screening, 89.53% for three-class classification-with ViT-small and DenseNet-121 exceeding their centralized counterparts in several configurations. Under strong DP constraints (noise multiplier η=2.0, yielding conservative privacy budget estimates of ε<1.0 with δ=10-5), screening accuracy remains above 82%, though diagnostic tasks incur substantial degradation (best 68.42%). Our findings provide empirical guidance on architecture selection, FL algorithm choice, and privacy-utility trade-offs for privacy-preserving breast cancer diagnosis, while identifying key challenges for clinical deployment.