Applying Federated Learning for Breast Cancer Prediction

Abstract

Breast cancer detection from histopathology images has become an important application of deep learning due to the need for early and accurate diagnosis. However, most conventional training pipelines rely on centralized datasets, where sensitive medical images must be stored in a single location. This raises significant concerns regarding patient confidentiality and compliance with regulations such as GDPR and HIPAA. To address these limitations, this work develops a privacy-preserving breast cancer classification framework using Federated Learning. The proposed system fine-tunes a VGG16-based Convolutional Neural Network, enabling multiple medical centers to collaboratively train a shared model without exposing raw patient data. Each participating client performs local training, while the central server aggregates model updates using the Federated Averaging (FedAvg) algorithm. The dataset was preprocessed by removing mask files, resizing images, applying augmentation, and organizing samples into training, validation, and test splits. Experimental evaluation demonstrates that the federated approach achieves reliable multi-class classification performance, reaching an accuracy of 91.5% while maintaining complete data privacy. These results indicate that federated training can serve as a practical alternative to centralized learning in real world medical environments.