Privacy-Preserving Breast Density Classification in Mammograms Using Fuzzy C-Means and Homomorphic Encryption

Abstract

Breast density is a significant risk factor for breast cancer, making its accurate assessment crucial for personalized screening strategies and improved patient outcomes. This study presents a privacy-preserving automated method for breast density classification using craniocaudal (CC) view mammograms from the VinDr-Mammo dataset. The approach combines Fuzzy C-Means clustering for tissue segmentation and morphological and texture feature extraction. Random Forest and XGBoost classifiers were trained on these features, achieving an accuracy of approximately 82 %. To address privacy concerns associated with sensitive medical data, the study employs Fully Homomorphic Encryption (FHE) via the Concrete-ml library. This allows encrypted models to perform secure computations directly on encrypted data, ensuring data privacy without sacrificing classification performance. Results demonstrate that privacy-preserving classifiers achieve accuracy comparable to traditional models, underscoring the feasibility of applying secure machine learning in clinical workflows. A practical workflow is also proposed to enable collaborative use between centralized hospitals and resource-limited health centers.