【摘要】 多中心疾病诊断方法通过整合不同医疗机构的样本信息到一台服务器上，集中训练来提高预测的准确性，有效解决了医疗领域小样本的问题.但仍存在两个问题：不同医疗机构的数据分布不同以及无法保护病人的隐私.基于此，设计了一种应用在多站点脑疾病诊断领域中隐私保护的联邦知识蒸馏算法.首先，设计了服务器端基于批标准化的加权平均算法，帮助联邦模型提取各个医疗机构数据分布无关的特征.之后，在客户端设计了联邦教师模型-本地学生模型的框架，部署了本地分类器，利用蒸馏损失保证模型提取本地化特征，利用分类损失保证模型性能稳定.实验结果表明，该算法在自闭症及精神分裂症数据集上均优于现有的其他算法.更多还原

【Abstract】 The multi-site disease diagnosis method can improve the accuracy of prediction by integrating the sample information of different medical institutions into one server, which effectively solves the problem of small sample size in the medical field. However, most of these approaches have two problems in the medical field which being the different distribution of data in different medical institutions and the inability to protect patient privacy. Based on these, we design a federal knowledge distillation algorithm for privacy protection in multi-site brain disease diagnosis. Firstly, a weighted average algorithm based on batch standardization is designed on the server to help the federated model to extract the distribution independent feature of each medical institution. Then, the framework of federated teacher model-local student model is designed on the client, and local classifier is deployed. The distillation loss guarantee model is used to extract localized features, and the classification loss is used to ensure the stable performance of the model. Experimental results show that the proposed algorithm is superior to other existing algorithms in autism and schizophrenia datasets.更多还原