陆森良1, 冯宝2, 徐坤财1, 陈业航2, 陈相猛3(1.桂林电子科技大学电子工程与自动化学院;2.桂林航天工业学院智能检测与信息处理实验室;3.江门市中心医院医学影像智能计算及应用实验室)
目的 针对联邦学习中多中心医学数据的异质性特征导致全局模型性能不佳的问题，提出一种基于特征迁移的自适应个性化联邦学习算法(Adaptive Personalized Federated learning via Feature Transfer, APFFT)。方法 首先,，为降低全局模型中异质性特征信息影响，提出鲁棒特征选择网络(Robust Feature Selection network, RFS-Net)构建个性化本地模型。RFS-Net通过学习两个迁移权重分别确定全局模型向本地模型迁移时的有效特征以及特征迁移的目的地，并构建基于迁移权重的迁移损失函数以加强本地模型对全局模型中有效特征的注意力，从而构建个性化本地模型。然后，为过滤各本地模型中异质性特征信息，利用自适应聚合网络(Adaptive Aggregation Network, AA-Net)聚合全局模型。AA-Net基于全局模型交叉熵变化更新迁移权重并构建聚合损失，从而使得各本地模型向全局模型迁移鲁棒特征，提高全局模型的特征表达能力。结果 在3种医学图像分类任务上与4种现有方法进行了比较实验，在肺结核肺腺癌分类任务中，各中心AUC 分别为0.7915,0.7981,0.76,0.7057,0.8069；在乳腺癌组织学图像分类任务中，各中心准确率分别为0.9849,0.9808,0.9835,0.9826,0.9834；在肺结节良恶性分类任务中,各中心AUC分别为 0.8097,0.8498,0.7848,0.7923。结论 所提出的联邦学习方法，降低了多中心的异质性特征影响，实现基于鲁棒特征的个性化本地模型自适应构建和全局模型自适应聚合，模型性能有较大提升。
Adaptive Personalized Federated learning: Transferring robust features to build medical image classification models
LUSenliang, FENGBao1, XU Kuncai2, CHEN Yehang1, CHEN Xiangmeng3(1.Laboratory of Intelligent Detection and Information Processing, Guilin University of Aerospace Technology;2.School of Electronic Engineering and Automation, Guilin University of Electronic Technology;3.Laboratory of Intelligent Computing and Application of Medical imaging, Jiangmen Central Hospital)
Objective Patient data cannot be shared among medical institutions due to medical data confidentiality regulations, which greatly limits the data scale. Federated learning ensures that all clients can train local models and aggregate global models in a decentralized way without sharing data. However, the heterogeneity of medical data greatly affects the aggregation and deployment of global models in federated learning. In most federated learning methods, the aggregation of global model parameters is achieved by multiplying the fixed weight with the local model parameters and summing them, and the local model personalization method requires a lot of manual experiments to select the appropriate model layer for personalization construction. Although these methods can realize the aggregation of global models or the construction of personalized local models, they cannot automatically aggregate global model parameters and construct personalized local models, and lack sufficient pertinence to the heterogeneity characteristics. Therefore, an adaptive personalized federated learning algorithm via feature transfer(APFFT) is proposed, which can automatically identify and select robust features for personalized local model construction and global model aggregation, and suppress and filter heterogeneous feature information. Method First, to construct a personalized local model, Robust Feature Selection network(RFS-Net) was proposed in this paper. RFS-Net can automatically identify and select features by calculating the transfer weights and the amount of feature transfer based on the model representation. When transferring features from global model to local model, RFS-Net constructs transfer loss functions based on transfer weights and the amount of feature transfer to constrain the local model and strengthen its attention to effective transfer features. In the Aggregation of global model, the Adaptive Aggregation Network (AA-Net) was proposed to transfer features from the local model to the global model. AA-Net updated the transfer weight and constructed the aggregation loss based on the cross-entropy change of the global model for filtering the heterogeneity feature information of each local model. In this paper, PyTorch is used to build and train the models, ResNet18 is used as Convolutional Neural Networks(CNN) structure , RFS-Net and AA-Net are composed of full connection layers, pooling layers, Softmax layers and Relu6 layers. The parameters of RFS-Net, AA-Net and CNN network were updated by stochastic gradient descent with momentum of 0.9. Experiments were carried out on three medical image data sets, namely the non-public data set of pulmonary adenocarcinoma and tuberculosis classification, the public data set Camelyon17 and the public data set LIDC. The data set of pulmonary adenocarcinoma tuberculosis classification came from 5 hospitals, with a total of 1009 cases, among which, Center 1(training set n=260, test set =242), Center 2(training set n=34, test set =54), Center 3(training set n=39, test set =40), Center 4(training set n=145, test set =108), Center 5(training set n=36, test set =51), in the experiment, the learning rate and decay rate of RFS-Net and AA-Net are both 0.0001, and the learning rate and decay rate of CNN network are 0.001 and 0.0005. Focalloss is used to calculate cross entropy. In addition, in the diagnosis of tuberculosis and lung adenocarcinoma, Gender, age, and nodule size in clinical information are of great reference value. Therefore, we made statistics on these information, and the results showed that in Center 2, the overall age and nodule size were small, while in center 4, the overall nodule size was large, which had a certain gap with the global average level. Camelyon17 was composed of 450,000 histological images from five hospitals. In the experiment, the learning rate and decay rate of CNN network, RFS-Net and AA-Net were all 0.0001. Standard cross entropy was used to constrain CNN network training. LIDC data came from 7 research institutions and 8 medical image companies, with a total of 1018 cases. Lesions with grade 1-2 malignancy were classified as benign, and those with grade 4-5 malignancy were classified as malignant. Finally, a total of 1746 lesions were included in the data set to simulate the federal learning application scenario, and were randomly divided into 4 centers according to the cases. Center 1(training set n=254, test set n=169), center 2(training set n=263, test set n=190), center 3(training set n=305, test set n=124), center 4(training set n=247, test set n=194), in the experiment, The learning rate and decay rate of RFS-Net and AA-Net are both 0.0001, and the learning rate and decay rate of CNN network are 0.001 and 0.0001. The cross entropy loss is calculated using standard cross entropy. Result Three kinds of medical image classification tasks were compared with four existing methods. The evaluation indexes included Receiver Operating Characteristic and accuracy. The experimental results show that in tuberculosis lung adenocarcinoma classification task, the center test set of end-to-end AUC were 0.7915, 0.7981, 0.76, 0.7057, 0.8069; In breast cancer histological image classification task, the center test set of end-to-end accuracy were 0.9849, 0.9808, 0.9835, 0.9826, 0.9834; In pulmonary nodules benign and malignancy classification task, the center test set of end-to-end AUC were 0.8097, 0.8498, 0.7848, 0.7923. Conclusion The federated learning method proposed in this paper can reduce the influence of heterogeneous characteristics and realize the adaptive construction of personalized local models and the adaptive aggregation of global models. The results show that our model is superior to several existing federated learning methods, and the model performance is greatly improved.