Sharma Shagun, Guleria Kalpna, Dogra Ayush
Chitkara University Institute of Engineering and Technology, Chitkara University, Punjab, India.
Curr Med Imaging. 2025;21:e15734056333970. doi: 10.2174/0115734056333970241212132150.
Pneumonia is an acute respiratory infection that has emerged as the predominant catalyst for escalating mortality rates worldwide. In the pursuit of the prevention and prediction of pneumonia, this work employs the development of an advanced deep-learning model by using a federated learning framework. The deep learning models rely on the utilization of a centralized system for disease prediction on the medical imaging data and pose risks of data breaches and exploitation; however, federated learning is a decentralized architecture which significantly reduces data privacy concerns.
The federated learning works in a distributed architecture by sending a global model to clients rather than sending the data to the model. The proposed federated deep learning-based FedPneu computer-aided diagnosis model has been implemented in 2, 3, 4, and 5 clients architecture for early pneumonia detection using X-ray images. The key parameters configuration include batch size, learning rate, optimizer, decay, momentum, epochs, rounds, and random-split as 32, 0.0001, SGD, 0.000001, 0.9, 10, 100, and 42, respectively.
The results of the proposed federated deep learning-based FedPneu model have been provided in terms of round-wise accuracy, loss, and computational time. The highest accuracy of 85.632% has been achieved with 2-clients federated deep learning architecture, whereas, 3, 4, and 5 clients architecture achieved 85.536%, 76.112%, and 74.123% accuracies, respectively.
In the proposed privacy-protected federated deep learning-based FedPneu model, the two-client architecture has been resulted as the most optimal framework for pneumonia detection among 3-clients, 4-clients, and 5-clients architecture. The model works in a collaborative and privacyprotected framework with a multi-silo dataset which could be highly beneficial for healthcare departments to maintain patient's data privacy with improved prediction outcomes.
肺炎是一种急性呼吸道感染，已成为全球死亡率上升的主要催化剂。为了预防和预测肺炎，本研究利用联邦学习框架开发了一种先进的深度学习模型。深度学习模型依赖于使用集中式系统对医学影像数据进行疾病预测，存在数据泄露和被利用的风险；然而，联邦学习是一种去中心化架构，可显著降低数据隐私问题。
联邦学习通过将全局模型发送给客户端而非将数据发送给模型，在分布式架构中运行。所提出的基于联邦深度学习的FedPneu计算机辅助诊断模型已在2、3、4和5客户端架构中实现，用于使用X射线图像进行早期肺炎检测。关键参数配置包括批量大小、学习率、优化器、衰减、动量、轮次、回合数和随机分割，分别为32、0.0001、随机梯度下降、0.000001、0.9、10、100和42。
所提出的基于联邦深度学习的FedPneu模型的结果已按回合准确率、损失和计算时间给出。2客户端联邦深度学习架构实现了最高准确率85.632%，而3、4和5客户端架构的准确率分别为85.536%、76.112%和74.123%。
在所提出的基于隐私保护的联邦深度学习的FedPneu模型中，两客户端架构是3客户端、4客户端和5客户端架构中肺炎检测的最优框架。该模型在具有多源数据集的协作和隐私保护框架中运行，这对于医疗部门维护患者数据隐私并改善预测结果可能非常有益。
