摘要
目的比较不同传统深度学习模式在肺癌诊断和分类中的应用价值。方法选取2016年1月至2017年11月在长沙市第一医院肿瘤内科接受治疗的33例患者为研究对象。获取非小细胞肺癌和小细胞肺癌活检标本,并进行染色。切片标本由2名经验丰富的病理学家进行诊断。采用多种深度学习方法区分癌症和非癌症活检。比较不同传统深度学习模式在肺癌诊断和分类中的应用价值。结果研究测试了几种流行的基于图像块分类的CNN架构:AlexNet、VGG、ResNet和SqueezeNet,比较两种类型的训练方案:从零开始训练和对整个预训练网络进行微调。深度学习模型AUC更合理(0.8810~0.9119),除ResNet-50外,从零开始训练的AUC高于对整个网络的微调。结论通过深度学习分析,可加快对全切片图像(WSI)的检测速度,且与病理学家保持相似的检出率。
Objective To compare different conventional deep learning models in lung cancer diagnosis and classification.Methods 33 patients who received treatment in the Department of Medical Oncology of First Hospital of Changsha City from January 2016 to November 2017 were selected as reseach subjects.The biopsy specimens of non-small cell lung cancer and small cell lung cancer were obtained and stained.The biopsy specimen was diagnosed by two experienced pathologists.Multiple deep learning methods were used to distinguish between cancer and non-cancer biopsy specimens.Compared the application value of different traditional deep learning models in the diagnosis and classification of lung cancer.Results This study tested several popular CNN architectures for the patch-based classification:AlexNet,VGG,ResNet and SqueezeNet,two types of training schemes were compared:training from scratch and fine-tuning the entire pre-training network.Deep learning models give reasonable AUC(0.8810-0.9119),training from scratch showed AUC higher than fine tuning the whole network except ResNet-50.Conclusion The deep learning analysis could speed up the detection process for the whole-slide image(WSI)and keep the comparable detection rate with human observer.
作者
李斌
李科宇
汤渝玲
李慧
LI Bin;LI Keyu;TANG Yuling;LI Hui(Changsha Hospital Affiliated to University of South China,Changsha,Hunan,410005,China;Department of Respiratory Medicine,the First Hospital of Changsha City,Changsha,Hunan,410005,China)
出处
《当代医学》
2021年第9期89-93,共5页
Contemporary Medicine
基金
湖南省卫生计生委科研计划课题项目(B20180393)
长沙市科技计划项目(kq1801133)
湖南省自然科学基金(2019JJ80111)。
关键词
深度学习
卷积神经网络
人工智能
肺癌
诊断
Deep learning
Convolutional Neural Networks
Artificial intelligence
Lung cancer
Diagnosis
