FTAD-Net: Federated Transfer-Adversarial Learning for Robust Breast Cancer Detection

Breast cancer remains the second leading cause of cancer-related mortality among women worldwide, making early and accurate diagnosis pivotal for improving patient survival rates. Although deep learning (DL) models have demonstrated remarkable potential in automated breast cancer detection, their large-scale clinical deployment is hindered by challenges such as limited annotated data, institutional data silos, and stringent privacy regulations. To overcome these limitations, this study proposes a Federated Transfer-Adversarial DenseNet (FTAD-Net), a state-of-the-art privacy-preserving and domain-generalizable DL framework designed for collaborative breast cancer diagnosis across multiple medical institu- tions. The proposed FTAD-Net introduces three major innovations. First, a federated transfer learning paradigm enables knowledge sharing among decentralized institutions without exchanging raw data, ensuring data confidentiality through secure gradient aggregation. Second, a domain-adversarial feature alignment module mitigates inter-institutional domain shifts by enforcing invariant feature representations across heterogeneous imaging modalities, including mammography and MRI. Third, a self-attentive feature refinement mechanism is integrated within the DenseNet backbone, enhancing discrimination of lesion-relevant regions while improving feature interpretability. The model was collaboratively trained on datasets from BreakHis, MIAS, and DDSM, encompassing diverse imaging conditions and patient populations. Ex- perimental evaluations demonstrate that FTAD-Net achieves a classification accuracy of 97.52%, surpassing existing centralized and federated approaches by up to 4.7%, with an average inference time of 10 seconds per case. Overall, the proposed framework provides a scalable, privacy-preserving, and domain-robust solution for early and reliable breast cancer detection.