Privacy-Preserving Breast Cancer Classification via Federated Learning on Multimodal Image Data

Abstract

Breast cancer significantly affects women’s health worldwide, underscoring the critical

need for advanced diagnostic approaches to improve early detection and treatment out- comes. This paper proposes a privacy-preserving framework for breast cancer classification

using federated learning integrated with multimodal imaging data, including mammogra- phy, ultrasound, and histopathology. It will enable FL to collaboratively study model

training across various institutions by not sharing sensitive information of patients, which

indeed may meet the requirements of GDPR, HIPAA, and related data-protection poli- cies. Herein, it proposes a hybrid encoder for classifying benign and malignant lesions by

combining the powers of CNNs with mechanisms of attention and feature fusion in a more

sophisticated way. The methodology incorporates extensive preprocessing and feature ex- traction with SMOTE to handle class imbalance. Validated on real-world multimodal

datasets within a federated framework, the model achieved superior performance with a test accuracy of 97.02% compared to traditional centralized approaches. Ablation study further optimized the model components relating to feature selection, pooling layers, and learning rates. The system is scalable, computationally efficient, and robust, offering huge reductions in false positives and negatives with data privacy. This research reflects the transformative potential of FL in healthcare, paving the way for an ethical, scalable, and effective diagnosis of breast cancer. Future work will incorporate explainable AI and a web-based platform for clinical decision-making in real-time, furthering the scope of this innovative approach.