FedBCFA: a boundary-constrained feature alignment federated learning framework for multi-center breast MRI classification

Abstract

Multi-center medical image modeling faces persistent challenges arising from data privacy constraints and cross-institutional distribution discrepancies. Federated learning provides a feasible paradigm for collaborative training of intelligent diagnostic models for breast magnetic resonance imaging (MRI) without sharing raw imaging data. However, under non-independent and identically distributed (non-IID) settings, variations in imaging protocols, cohort composition, and class distributions across institutions may induce representation shifts and class-boundary ambiguity, particularly compromising the stable discrimination between benign and malignant lesions. To address these challenges, this paper proposes FedBCFA, a Federated Boundary-Constrained Feature Alignment method for three-class classification of multi-center breast MRI, including no-lesion, benign, and malignant cases. During federated training, FedBCFA integrates class-prototype feature alignment with a boundary-constrained learning mechanism to alleviate inter-client representation inconsistency and promote stable learning of discriminative class boundaries. In addition, benign-oriented frequency-domain augmentation and a constrained model selection strategy are introduced to improve the recognition of minority and boundary samples. Experimental results demonstrate that the proposed method achieves favorable overall performance in multi-center breast MRI classification, effectively mitigates class confusion under non-IID conditions, and improves the discriminative stability for benign and malignant lesions. Overall, FedBCFA enhances the generalization capability of multi-center breast MRI classification models under privacy-preserving constraints and provides a feasible solution for cross-institutional computer-aided diagnosis in medical imaging.