计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (23): 228-236.DOI: 10.3778/j.issn.1002-8331.2208-0207

• 图形图像处理 • 上一篇    下一篇

多模式特征融合网络肺结节良恶性分类方法

尹智贤,夏克文,武盼盼   

  1. 1.河北工业大学 电子信息工程学院,天津 300401
    2.天津师范大学 计算机与信息工程学院,天津 300387
  • 出版日期:2023-12-01 发布日期:2023-12-01

Classification of Benign and Malignant Pulmonary Nodules by Multimodal Feature Fusion Network

YIN Zhixian, XIA Kewen, WU Panpan   

  1. 1.School of Electronics and Information Engineering, Hebei University of Technology, Tianjin 300401, China
    2.School of Computer and Information Engineering, Tianjin Normal University, Tianjin 300387, China
  • Online:2023-12-01 Published:2023-12-01

摘要: 胸部计算机断层扫描(computed tomography,CT)中肺结节良恶性的精确分类对于肺癌的早期诊断具有重要意义。然而,CT影像中肺结节背景的复杂,以及图像特征提取不全面等问题,为实现肺结节良恶性的精确分类带来困扰。为此,提出了多模式特征融合网络肺结节良恶性分类方法。具体地,以MobileNet V3为骨干网络,以原始肺结节CT图像及提取出的结节图像为输入,设计了一种双路径特征提取网络,不仅能够有效提取原CT图像的全局信息,还能有效挖掘肺结节区域的判别性特征,以弥补结节较小时网络过多关注其周围组织从而产生误判的问题。此外,在特征提取阶段引入convolutional block attention module(CBAM)和通道混洗机制,进一步增强了网络的特征表达能力。同时,对原MobileNet V3网络结构做出修改,删除最后四组基于倒残差结构的bottlenecks(bnecks)模块,使模型能够以较小的时间和空间复杂度精确诊断恶性结节。在LIDC-IDRI数据集上的实验表明,提出的方法能够在显著降低网络参数量和FLOPs的同时实现对肺结节良恶性的精确分类,分类准确率、敏感性、特异性、精确率、F1值和AUC值分别达到了93.71%、94.03%、93.48%、95.56%、92.65%和98.66%。

关键词: 肺结节良恶性分类, 特征融合, 卷积块注意力模块(CBAM), 通道混洗, MobileNet V3

Abstract: Accurate classification of benign and malignant pulmonary nodules in computed tomography(CT) is crucial for the early diagnosis of lung cancer. However, due to the complexity of the background of pulmonary nodules in CT images and the incompleteness of image feature extraction, it is difficult to realize the accurate classification of benign and malignant pulmonary nodules. To this end, a method of classification of benign and malignant pulmonary nodules by multimodal feature fusion network is proposed. Specifically, with MobileNet V3 as the backbone network and the original CT image of the lung nodule and the extracted nodule image as the input, a dual-path feature extraction network is designed, which can not only effectively extract the global information of the original CT image, but also the discriminative features of the lung nodule area are effectively mined to make up for the problem that when the nodule is small, the network pays too much attention to the surrounding tissue, resulting in misjudgment. For further enhancing feature expression ability of the network, convolutional block attention module(CBAM) and channel shuffle mechanism are employed. Meanwhile, the last four bottlenecks(bnecks) based on inverted residuals in the original MobileNet V3 network are deleted, so that malignant nodules can be accurately diagnosed with less time and space complexity. The experiments on LIDC-IDRI dataset show that this method can significantly reduce the network parameters and FLOPs, and realize the accurate classification of benign and malignant pulmonary nodules. The classification accuracy, sensitivity, specificity, precision, F1 score and AUC are 93.71%, 94.03%, 93.48%, 95.56%, 92.65% and 98.66% respectively.

Key words: classification of benign and malignant pulmonary nodules, feature fusion, convolutional block attention module(CBAM), channel shuffle, MobileNet V3