Journal of Frontiers of Computer Science and Technology ›› 2024, Vol. 18 ›› Issue (7): 1705-1724.DOI: 10.3778/j.issn.1673-9418.2310064

• Frontiers·Surveys • Previous Articles     Next Articles

Review of Classification Methods for Lung Nodules in CT Images

LI Jiancheng, CAO Lu, HE Xiquan, LIAO Junhong   

  1. 1. School of Electronic and Information Engineering, Wuyi University, Jiangmen, Guangdong 529020, China
    2. Department of Respiratory and Critical Care Medicine, The People’s Hospital of Jiangmen, Jiangmen, Guangdong 529000, China
  • Online:2024-07-01 Published:2024-06-28

CT影像下的肺结节分类方法研究综述

利建铖,曹路,何锡权,廖军红   

  1. 1. 五邑大学 电子与信息工程学院,广东 江门 529020
    2. 江门市人民医院 呼吸与危重症医学科,广东 江门 529000

Abstract: In recent years, deep learning has been widely applied to various classification tasks due to its capability in automatically extracting features and superior classification performance. Research on the classification of lung nodules has gradually shifted from traditional methods that involve manual feature extraction to deep learning-based classification approaches. To better investigate the benign-malignant classification of lung nodules in CT images, the current status of deep learning methods based on convolutional neural network (CNN) in the research of benign and malignant classification of lung nodules is summarized. Firstly, this paper introduces commonly used publicly available datasets for lung nodule classification, including their contents, limitations, and download sources. Secondly, it outlines commonly used performance evaluation metrics. It then highlights the recent research work on deep learning methods for lung nodule classification: current methods for lung nodule classification are categorized as using only CNN, introducing an attention mechanism in CNN, multi-view learning, multi-modal learning, and using migration learning, adversarial neural networks, and other methods, repectively, from the level of network structure and data. Meanwhile, this paper further summarizes the network structures, advantages and disadvantages of these classification methods. A comparative analysis is conducted on the benign-malignant classification performance of lung nodule classification methods based on these aspects over the past three years using publicly available nodule datasets. Finally, this paper discusses current challenges and explores further research directions in the field of lung nodule classification.

Key words: deep learning, lung nodule classification, convolutional neural networks (CNN), feature extraction

摘要: 近年来,深度学习因其具有自动提取特征的能力以及更好的分类性能而被广泛应用于各种分类任务之中。肺结节的分类研究也逐渐从手工提取特征的传统方法向基于深度学习的分类方法转变。为了更好地对CT影像下的肺结节进行良恶性分类研究,对以卷积神经网络(CNN)为主的深度学习方法在肺结节良恶性分类研究的现状进行梳理和归纳总结。首先介绍了目前常用的肺结节公开数据集,包括其内容、局限性以及下载地址。其次总结了常用的性能评价指标。然后重点介绍了近年来深度学习方法在肺结节分类中的研究工作:分别从网络结构层面和数据层面将当前肺结节分类方法归类为仅使用卷积神经网络、在卷积神经网络中引入注意力机制、多视图学习、多模态学习以及使用迁移学习、对抗神经网络这些方法;同时总结了这些分类方法的网络结构以及优缺点,并且对比了近三年的基于这些内容的肺结节分类方法在肺结节公开数据上的良恶性分类表现。最后讨论了目前肺结节分类中存在的问题并探索进一步的研究方向。

关键词: 深度学习, 肺结节分类, 卷积神经网络(CNN), 特征提取