计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (23): 167-174.DOI: 10.3778/j.issn.1002-8331.2008-0120

• 模式识别与人工智能 • 上一篇    下一篇

融合注意力的多维多特征浸润性腺癌诊断

鲍涧颖,张岩,徐建林,莫锦秋   

  1. 1.上海交通大学 机械与动力工程学院,上海 200240
    2.上海交通大学附属胸科医院 呼吸内科,上海 200030
  • 出版日期:2020-12-01 发布日期:2020-11-30

Diagnosis of Adenocarcinoma?Based on Multi-dimensional Features and Attention Fusion

BAO Jianying, ZHANG Yan, XU Jianlin, MO Jinqiu   

  1. 1.School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China
    2.Department of Pulmonary, Shanghai Chest Hospital, Shanghai Jiao Tong University, Shanghai 200030, China
  • Online:2020-12-01 Published:2020-11-30

摘要:

目前早期肺癌筛查主要依据影像科医生的经验由肺部CT影像来诊断肺部结节的良恶性,腺癌中愈后最差的浸润性腺癌结节需通过术中快速冰冻病理进行诊断,且快速冰冻病理对小直径结节诊断准确率较低。针对上述问题,提出了依据CT影像对磨玻璃肺小结节中的浸润性腺癌结节进行诊断的算法。根据结节空间信息及平面特征设计不同维数的样本数据,即3D空间样本与2D平面特征样本,并基于注意力机制及残差学习单元设计网络结构,搭建2D及3D神经网络。通过融合不同维数的网络提取的特征向量,最终得到浸润性腺癌结节的诊断结果。该算法在采集自上海胸科医院的有手术病理结果的结节直径为5~20?mm的1?760份磨玻璃小结节样本上进行研究,其中浸润性腺癌的结节样本共340份,非浸润性的结节样本共1?420份,在该实例数据集上通过交叉验证,该算法的分类准确率为82.7%,敏感度为82.9%,特异度为82.6%。

关键词: 注意力, 多维度, 多特征, 特征融合, 浸润性腺癌诊断

Abstract:

Nowadays, lung cancer is one of cancers that endanger human health and have a high fatality rate. Annual CT screening can detect curable pulmonary nodules. At present, it mainly depends on the experience of radiologists to diagnose pulmonary nodules. The adenocarcinoma nodules which have the worst infiltrating should be diagnosed by the method of intraoperative rapid freezing, and the accuracy for small diameter nodules is low. For above problems, this paper proposes an algorithm for the diagnosis of adenocarcinoma nodules in small glass-ground lung nodules based on CT images. Two different dimensions of data are generated for extracting spatial and plane characteristics, and two different dimensions of networks are constructed by the residual learning unit combined with the attention mechanism. By combining the feature vectors extracted by two networks, the diagnosis of adenocarcinoma nodules can be concluded. This algorithm is experimented in the dataset collected from Shanghai Chest Hospital. It consists of 1,760 small glass-ground lung nodules including 340 adenocarcinoma nodules and 1,420 other nodules. The cross validation is performed in this experiment, the classification accuracy of this algorithm reaches 82.7%, the sensitivity reaches 82.9% and the specificity reaches 82.6%.

Key words: attention, multi-dimensional, multi-features, feature fusion, diagnosis of adenocarcinoma