计算机应用 ›› 2020, Vol. 40 ›› Issue (9): 2571-2576.DOI: 10.11772/j.issn.1001-9081.2019122122

• 人工智能 • 上一篇    下一篇

基于特征金字塔网络的肺结节检测

高智勇, 黄金镇, 杜程刚   

  1. 中南民族大学 生物医学工程学院, 武汉 430074
  • 收稿日期:2019-12-19 修回日期:2020-05-18 出版日期:2020-09-10 发布日期:2020-07-08
  • 通讯作者: 高智勇
  • 作者简介:高智勇(1972-),男,湖北浠水人,副教授,博士,主要研究方向:图像分割、图像识别、认知计算;黄金镇(1994-),男,广西河池人,硕士研究生,主要研究方向:图像分割、图像识别、目标检测;杜程刚(1995-),男,湖北武汉人,硕士研究生,主要研究方向:图像分割、图像识别、目标检测。
  • 基金资助:
    国家自然科学基金资助项目(61240059);中央高校基本科研业务费专项(CZY20007)。

Pulmonary nodule detection based on feature pyramid networks

GAO Zhiyong, HUANG Jinzhen, DU Chenggang   

  1. College of Biomedical Engineering, South-Central University for Nationalities, Wuhan Hubei 430074, China
  • Received:2019-12-19 Revised:2020-05-18 Online:2020-09-10 Published:2020-07-08
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (61240059), the Fundamental Research Funds for the Central Universities (CZY20007).

摘要: 针对计算机断层扫描(CT)影像中肺结节尺寸变化较大、尺寸小且不规则等特点导致的检测敏感度较低的问题,提出了基于特征金字塔网络(FPN)的肺结节检测方法。首先,利用FPN提取结节的多尺度特征,并强化小目标及目标边界细节的特征;其次,在FPN的基础上设计语义分割网络(名为掩模特征金字塔网络(Mask FPN))用于快速准确地分割提取肺实质,作为目标候选区域定位图像;并且,在FPN顶层添加反卷积层,采用多尺度预测策略改进快速区域卷积神经网络(Faster R-CNN)以提高检测性能;最后,针对肺结节数据集的正负样本不平衡问题,在区域候选网络(RPN)模块采用焦点损失函数以提高结节的检出率。所提方法在公开数据集LUNA16上进行实验,结果表明,利用FPN和反卷积层改进的新网络对结节检测效果有一定的帮助,采用焦点损失函数也有一定效果。综合多种改进,当平均每个扫描件的候选结节数为46.7时,所提方法的肺结节检测敏感度指标为95.7%,与其他卷积神经网络方法如Faster R-CNN、UNet等相比,具有较高的敏感性。所提方法能够较好地提取不同尺度上的结节特征,提高CT图像肺结节检测的敏感度,同时对于较小的结节也能有效检测,能更有效地辅助肺癌的诊断治疗。

关键词: 肺结节检测, 肺实质分割, 特征金字塔网络, 卷积神经网络, 多尺度

Abstract: Pulmonary nodules in Computerized Tomography (CT) images have large size variation as well as small and irregular size which leads to low detection sensitivity. In order to solve this problem, a method based on Feature Pyramid Network (FPN) was proposed. First, FPN was used to extract multi-scale features of nodules and strengthen the features of small objects and object boundary details. Second, a semantic segmentation network (named Mask FPN) was designed based on the FPN to segment and extract the pulmonary parenchyma quickly and accurately, and the pulmonary parenchyma area could be used as location map of object proposals. At the same time, a deconvolution layer was added on the top layer of FPN and a multi-scale prediction strategy was used to optimize the Faster Region Convolution Neural Network (R-CNN) in order to improve the performance of pulmonary nodule detection. Finally, to solve the problem of imbalance of positive and negative samples in the pulmonary nodule dataset, Focal Loss function was used in the Region Proposed Network (RPN) module in order to increase the detection rate of nodules. The proposed algorithm was tested on the public dataset LUNA16. Experimental results show that the improved network with FPN and deconvolution layer is helpful to the detection of pulmonary nodules, and focal loss function is also helpful to the detection. Combining with multiple improvements, when the average number of candidate nodules per scan was 46.7, the sensitivity of the presented method was 95.7%, which indicates that the method is more sensitive than the other convolutional networks such as Faster Region-Convolutional Neural Network (Faster R-CNN) and UNet. The proposed method can extract nodule features of different scales effectively and improve the detection sensitivity of pulmonary nodules in CT images. Meantime, the method can also detect small nodules effectively, which is beneficial to the diagnosis and treatment of lung cancer.

Key words: pulmonary nodule detection, lung segmentation, Feature Pyramid Network (FPN), convolutional neural network, multi-scale

中图分类号: