基于自适应聚合权重联邦学习的肺结节CT图像分类  

Lung Nodule CT Image Classification Based on Adaptive Aggregate Weight Federated Learning

作　　者：侍江烽 冯宝[2] 陈业航[2] 陈相猛[3] Shi Jiangfeng;Feng Bao;Chen Yehang;Chen Xiangmeng(School of Electronic Engineering and Automation,Guilin University of Electronic Technology,Guilin 541004,Guangxi,China;Laboratory of Artificial Intelligence of Biomedicine,Guilin University of Aerospace Technology,Guilin 541004,Guangxi,China;Laboratory of Intelligent Computing and Application of Medical Imaging,Jiangmen Central Hospital,Jiangmen 529030,Guangdong,China)

机构地区：[1]桂林电子科技大学电子工程与自动化学院,广西桂林541004 [2]桂林航天工业学院生物医学与人工智能实验室,广西桂林541004 [3]江门市中心医院医学影像智能计算及应用实验室,广东江门529030

出　　处：《激光与光电子学进展》2023年第22期78-88,共11页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(81960324,62176104);广西自然科学基金(粤桂联合基金)(2021GXNSFAA075037);广东省医学科学技术研究基金(A2021138);桂林航天工业学院校级科研基金(XJ21KT24)。

摘　　要：针对目前医学影像面临多中心数据存在数据孤岛以及非独立同分布的问题(Non-IID),提出了一种基于自适应聚合权重的联邦学习算法(FedAaw)。在全局模型聚合过程中,提出准确率阈值来筛选出本地模型,并由中心服务器采用筛选后模型的准确率计算相应的聚合权重,从而对全局模型进行聚合,使得分类性能较佳的模型参与全局模型的构建,以达到缓解多中心数据Non-IID的问题。同时,为提高模型挖掘图像长短距离信息之间的能力,在本地和全局模型中引入多头自注意力(MHSA)机制。此外,为缓解端对端的冗余特征造成的模型过拟合问题,提取全局模型中卷积核的特征,并采用基于L1范数的稀疏贝叶斯极限学习机(SBELML_(1))的集成学习方法完成各中心数据的特征分类。最后,通过多次打乱不同中心的数据分布来验证FedAaw算法的抗干扰能力。5个中心的测试集AUC变化范围为中心1(0.7947~0.8037)、中心2(0.8105~0.8405)、中心3(0.6768~0.7758)、中心4(0.8496~0.9063)、中心5(0.8913~0.9348),该结果表明:FedAaw在多中心数据上具有良好的分类性能且抗干扰能力较强。The field of medical imaging currently faces the problems of data island and non-independent and independently distributed(Non-IID)variables in multi-center data.In this study,a federated learning algorithm based on adaptive aggregate weight(FedAaw)is proposed.Using a global model polymerization process,this study utilized the accuracy threshold to filter out the local model;the model accuracy is calculated by the center server.The corresponding weights of polymerization,which are updated in the global model,yielded models with better classification performances that are used to construct a global model,which helps address the problems associated with Non-IID multicenter data.Furthermore,to improve the applicability of the model to mining the information between the long and short distance of the image,the multi head self-attention mechanism is introduced to the local and global models.In addition,to address the problem of model overfitting caused by end-to-end redundant features,the convolution kernel features in the global model are extracted.The learning of sparse Bayesian extreme learning machine based on L1 norm(SBELML_(1))framework is used for the feature classification of the data obtained from each center.Finally,the anti-interference ability of the FedAaw algorithm is verified by shuffling the data distribution of different centers several times.The AUC ranges of the test sets used in the five centers are as follows:center 1:(0.7947‒0.8037),center 2:(0.8105‒0.8405),center 3:(0.6768‒0.7758),center 4:(0.8496‒0.9063),and center 5:(0.8913‒0.9348).These results indicate that FedAaw has good classification performance on multi-center data and a strong anti-interference ability.

关 键 词：自适应聚合权重 联邦学习 多头自注意力 L1范数的极限学习机 对抗验证 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391.41[电子电信—信息与通信工程]