基于集成知识蒸馏的肺病分析
Lung disease analysis using ensemble knowledge distillation
摘要点击 157  全文点击 29  投稿时间:2020-04-30  修订日期:2020-08-13
查看全文  查看/发表评论  下载PDF阅读器
DOI编号  10.7641/CTA.2020.00235
  2021,38(1):130-136
中文关键词  肺疾病  卷积神经网络  集成知识蒸馏  教师模型  学生模型
英文关键词  lung diseases  convolutional neural network  ensemble knowledge distillation  teacher model  student model
基金项目  国家自然科学基金项目(U1713212, 61836005, 61702341), 广东省杰出青年基金项目(2019B151502018), 深圳市技术研究基金项目(JSGG201805- 07182904693), 深圳市公共技术平台基金项目(GGFW2018021118145859).
作者单位E-mail
李坚强 深圳大学 lijq@szu.edu.cn 
王成 深圳大学  
黄志超 深圳大学  
陈杰 深圳大学 chenjie@szu.edu.cn 
中文摘要
      新冠肺炎以来, 肺部疾病引起了人们更大的关注. 肺音的特征与诊断是肺病理学中重要的组成部分. 现有 的肺音分析工作主要是对肺音的类型进行分类, 分类肺病的研究较少. 另外, 单个分类模型无法在保护隐私的前提 下融合多方数据, 复杂的模型也难以保证分类的实时性. 针对这些问题, 本文提出集成知识蒸馏的肺疾病分类模型. 首先从肺音音频中提取梅尔频谱特征, 然后建立多个二分类卷积神经网络模型作为教师模型, 最后通过集成知识蒸 馏技术, 将多个教师模型的知识集成到一个精简的多分类学生模型上. 实验表明, 该模型能够在预测准确率达95% 的情况下, 参数量比教师模型减少79%, 预测速度提升20%. 在同等条件下, 时耗仅有MobileNet–v3–small模型的6%, 实现实时性肺疾病分析.
英文摘要
      Since the outbreak of COVID–19, lung diseases have attracted more attention. The characteristics and diagnosis of lung sounds become an important part of pulmonary pathology. The existing works for lung sound analysis mainly aim to classify the types of abnormal lung sounds. There are few studies focusing on the classification of lung diseases. Moreover, a single classification model cannot take advantage of the train data from multiple sources due to privacy leakage concerns, and it is difficult for complex models to classify in real time. Therefore, this paper proposes a model for classifying lung diseases based on ensemble knowledge distillation. Firstly, Mel-spectrum features were extracted from lung sounds, and then multiple binary convolutional neural network models were established as teacher models. Finally, a simplified multi-class student model will learn the knowledge of multiple teacher models through the technology of ensemble knowledge distillation. Our experiments show that the student model reduces 79% of the parameters and improves the prediction efficiency by 20% than teacher model while achieving a predictive accuracy of 95%. Under the same condition, the student model only incurs 6% of the time that is used by the state-of-the-art MobileNet–v3–small model. Thus, our model has potential to be deployed in real world for real-time diagnosis of the lung diseases.