SFLFEM: A frequency-enhanced mamba framework for selective personalized federated breast ultrasound diagnosis.
Authors
Affiliations (4)
Affiliations (4)
- Engineering Department, Affiliated Hospital of Nanjing University of Chinese Medicine (Jiangsu Province Hospital of Chinese Medicine), Nanjing, 210029, Jiangsu, China.
- School of Biomedical Engineering, Faculty of Engineering, University of New South Wales, Sydney, NSW, 2052, Australia; ARC Centre of Excellence for Nanoscale Biophotonics, University of New South Wales, 11 Sydney, NSW, 2052, Australia.
- Engineering Department, Affiliated Hospital of Nanjing University of Chinese Medicine (Jiangsu Province Hospital of Chinese Medicine), Nanjing, 210029, Jiangsu, China. Electronic address: [email protected].
- Engineering Department, Affiliated Hospital of Nanjing University of Chinese Medicine (Jiangsu Province Hospital of Chinese Medicine), Nanjing, 210029, Jiangsu, China. Electronic address: [email protected].
Abstract
Breast ultrasound is essential for lesion assessment and early cancer screening. However, deploying robust deep learning models in clinical practice faces two primary challenges. The first is intrinsic: lesions often exhibit complex textures, speckle-corrupted appearances, and indistinct boundaries. The second is extrinsic: the data is distributed across isolated medical centers with significant heterogeneity. To address these coupled challenges, this study introduces SFLFEM, a unified framework that combines advanced feature representation with personalized federated optimization for privacy-preserving multi-center breast ultrasound diagnosis. FEMamba, a dual-domain backbone, integrates the long-range modeling capacity of State-Space Models with a wavelet-based frequency pathway. This architecture captures global contextual dependencies as well as frequency-sensitive boundary and textural cues. Both are critical for distinguishing challenging lesions. At the optimization level, the pFedBM algorithm introduces a Bidirectional Gradient Masking mechanism to mitigate client heterogeneity and cross-site distribution shift in federated learning. This approach adaptively separates model parameters into globally shared and locally preserved components. Locally sensitive shared parameters are promoted into personalized subspaces, and weakly client-specific parameters are demoted back to the shared global subspace. This design reduces negative transfer and improves site-specific adaptation. Comprehensive evaluations on four datasets demonstrate that SFLFEM achieves strong overall performance, with an average accuracy of 85.14%, AUC of 92.35%, and MCC of 70.62%. It achieves competitive performance against representative centralized classification backbones and consistently improves over the evaluated federated learning baselines. The largest gains of pFedBM were observed under heterogeneous and data-limited settings.