计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (10): 133-138.DOI: 10.3778/j.issn.1002-8331.2003-0012

• 模式识别与人工智能 • 上一篇    下一篇

基于三维各向异性卷积的肺结节良恶性分类

孙浩天,袁刚,杨杨,刘含秋,郑健,杨晓冬,张寅   

  1. 1.中国科学技术大学,合肥 230026
    2.中国科学院 苏州生物医学工程技术研究所 医学影像实验室,江苏 苏州 215163
    3.复旦大学附属华山医院 放射科,上海 200041
  • 出版日期:2021-05-15 发布日期:2021-05-10

3D Anisotropic Convolution Based Pulmonary Nodule Classification

SUN Haotian, YUAN Gang, YANG Yang, LIU Hanqiu, ZHENG Jian, YANG Xiaodong, ZHANG Yin   

  1. 1.University of Science and Technology of China, Hefei 230026, China
    2.Department of Medical Imaging, Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou, Jiangsu 215163, China
    3.Department of Radiology, Huashan Hospital, Fudan University, Shanghai 200041, China
  • Online:2021-05-15 Published:2021-05-10

摘要:

计算机断层扫描(Computer Tomography,CT)图像中肺结节的良恶性诊断对治疗方案的选择有非常重要的作用。目前基于深度学习的CT图像肺结节良恶性分类算法的一个研究趋势是充分利用CT图像的三维信息来设计网络,但由于不同CT设备采集的图像参数不同,不同样本的CT图像其层内及层间分辨率也多不相同,进行特征提取前需要进行额外的预处理工作。大多数文献的做法是采用插值的方法统一分辨率,然而这种方法会造成图像分辨率降低或计算量增加等问题。针对这一问题,提出了一种基于三维各向异性卷积的肺结节良恶性分类网络,通过将标准三维卷积拆分为[k×k×1]和[1×1×k]的两种三维各向异性卷积,避免了直接将三维卷积作用到原始CT图像上,从而避免了图像分辨率不同的影响。还提出了裁剪-非局部池化模块,通过中心裁剪和非局部池化操作,强化网络对结节区域的特征提取,同时使浅层网络也可以获取全局信息。在Lung Image Database Consortium and Image Database Resource Initiative(LIDC-IDRI)数据集上的实验表明,提出的三维各向异性卷积结合裁剪-非局部池化操作的神经网络能显著减少网络参数量,提升网络提取特征的能力,实现对肺结节良恶性的准确分类,分类的准确率、敏感性、特异性分别为91.53%、88.89%和93.27%,取得了比较好的分类性能。

关键词: 卷积神经网络, 各向异性卷积, 肺结节分类, 计算机辅助诊断, 计算机断层成像(CT)

Abstract:

The diagnosis of pulmonary nodules in CT images plays an important role in the selection of therapeutic plan. the research tendency of deep learning based classification algorithm is to make full use of 3D information of CT images, Due to the differences in CT scanners and scanning protocols, different CT samples have different in-plane resolutions on slice thicknesses, which need additional preprocessing before feature extraction. Most of the literatures adopt interpolation method to unify the resolution, but this method will cause the image resolution reduction or the increased computation burden. To solve this problem, this paper proposes a 3D anisotropic convolution based neural network for pulmonary nodules classification. By dividing the standard 3D convolution into two kinds of 3D anisotropic convolutions of [k×k×1] and [1×1×k], it avoids directly applying 3D convolution to the original CT images and will not be affected by different image resolutions. This paper also proposes a crop-non-local pooling module to enable the shallow network to obtain global information. Experiments on Lung Image Database Consortium and Image Database Resource Initiative(LIDC-IDRI) dataset show that the proposed architecture can significantly reduce the parameters of network, improve the ability of extracting features. The accuracy, sensitivity and specificity of pulmonary nodules classification are 91.53%, 88.89% and 93.27% respectively.

Key words: convolutional neural networks, anisotropic convolution, pulmonary nodule classification, computer aided diagnosis systems, Computed Tomography(CT)