Federated Deep Learning for Privacy-Preserving Breast Cancer Diagnosis from Multimodal Medical Data

Abstract

Breast cancer is one of the major causes of death among women population and an effective clinical intervention requires accurate diagnosis at the early stage of the disease. While recent progress in deep learning has been made on automated breast cancer analysis from mammographic images, centralized training still has some significant issues regarding patient privacy, institutional data sharing constraints, and the heterogeneity of medical data distributions. Current CNN and federated learning approaches also suffer from the inability to capture relationship between patients and similarity aware diagnostic patterns that leads to lower classification robustness and high communication. In the view of the gaps in existing work the Graph-Based Federated Breast Cancer Prediction Network (GFBC-Net) is proposed in this paper using the CBIS-DDSM mammography dataset. We design the proposed framework that combines the convolutional feature extraction and patient similarity modeling using a graph neural network under a federated learning setup. First, mammogram images are preprocessed, region of interest (ROI) is extracted, and the images are normalized, and deep features are generated using CNN. Next, patient similarity graphs are created based on cosine similarity and K-nearest neighbor algorithms, with patient cases as nodes in the graphs and feature-level relationships as weighted edges. The local graph models are trained in the different distributed clients of hospitals but secure federated aggregation updates the global model without transferring any raw medical images. The experimental results show that GFBC-Net was able to attain an accuracy of 94.27%, precision of 93.88%, recall of 93.21%, F1-score of 93.54%, and AUC of 96.12%, outperforming the conventional CNN, federated CNN and graph-based baseline models. The study shows that the graph-guided federated learning method can better assist in breast cancer classification, while ensuring classification reliability and institutional data security to protect privacy.