Federated Learning in Breast Cancer Diagnosis

Abstract

Breast cancer, which accounts for 29.7% of cancer cases among women and has a mortality rate of approximately 16.5%, it is the most common cancer in Brazil. To enhance early detection and minimize the occurrence of advanced cases, pathologists are increasingly utilizing deep learning techniques to analyze histopathological images. This study investigates and compares the performance of two convolutional neural networks, AlexNet and ResNet50, in federated and centralized training settings. By employing the BreakHis dataset, the research explores the effectiveness of federated learning in securely distributing model training across multiple clients and contrasts it with traditional centralized training approaches. This study investigates whether federated learning can achieve performance close to that of centralized approaches while ensuring privacy for breast cancer histopathology image classification. It is hypothesized that federated learning can provide comparable results to centralized approaches while ensuring the confidentiality of patient data. Our results confirm that federated learning performs similarly to centralized learning with the added benefits of maintaining privacy and supporting collaboration among multiple clients. The findings aim to advance the state of the art in breast cancer detection, addressing both the challenges of privacy and data distribution and the impact of different training methodologies on model performance.