Federated Learning Architecture for 3D Breast Cancer Image Classification

TL;DR

This paper introduces a privacy-preserving method for breast cancer detection using 3D images and federated learning, allowing hospitals to collaborate without sharing patient data.

Contribution

A novel federated learning architecture for 3D breast cancer image classification that preserves privacy while improving diagnostic accuracy.

Findings

The CNN model achieved 97.30% accuracy, which improved to 97.37% with federated learning.

Transfer learning models and autoencoders showed lower accuracy, ranging from 48.83% to 89.24%.

The CNN-FL framework balances diagnostic accuracy and data security effectively.

Abstract

Breast cancer is one of the most common and deadly diseases affecting women. Detecting it early can save many lives, but developing accurate computer systems for diagnosis usually requires sharing large amounts of patient data, which raises privacy concerns. In this study, we introduce a new method that allows hospitals to work together to improve breast cancer detection without sharing any sensitive data. Instead of sending patient images to a central location, each hospital trains its own model locally and shares only the learned information. These updates are then combined to create a stronger, global model. Our approach focuses on advanced three-dimensional breast images, which provide more detailed information for diagnosis. This work could help medical institutions collaborate securely and develop powerful, privacy-preserving tools to improve early detection and treatment of breast cancer.

Backgrouds: Breast cancer remains a major global health challenge, with early diagnosis playing a crucial role in improving patient survival rates. Among the available diagnostic techniques, mammography is widely employed for early detection. However, its effectiveness is often constrained by the complexity of image interpretation, which makes automated detection methods increasingly vital. Methods: In this study, we propose an advanced approach that leverages 3D mammographic imaging and integrates Federated Learning (FL) to enable decentralized, privacy-preserving model training across multiple institutions. To evaluate the effectiveness of this approach, we assess various machine learning models, including Convolutional Neural Networks (CNNs), Transfer Learning architectures (VGG16, VGG19, ResNet50), and AutoEncoders (AEs), using 3D mammographic data. Results: Our results indicate that the CNN model achieves an accuracy of 97.30%, which improves slightly to 97.37% when the model is combined with Federated Learning, highlighting both the predictive performance and privacy-preserving advantages of our method. In contrast, Transfer Learning models and AutoEncoders exhibit lower accuracies that range from 48.83% to 89.24%, revealing their limitations in the context of this specific task. Conclusions: These findings underscore the effectiveness of the CNN-FL framework as a robust tool for breast cancer detection, showing that this approach offers a promising balance between diagnostic accuracy and data security—two critical factors in medical imaging.

Figures