Journal of Computer Applications ›› 2020, Vol. 40 ›› Issue (7): 2110-2116.DOI: 10.11772/j.issn.1001-9081.2019122095

• Virtual reality and multimedia computing • Previous Articles     Next Articles

Detection method of pulmonary nodules based on improved residual structure

SHI Lukui1, MA Hongqi1, ZHANG Chaozong2, FAN Shiyan1   

  1. 1. School of Artificial Intelligence, Hebei University of Technology, Tianjin 300401, China;
    2. Hebei Science and Technology Information Institute, Shijiazhuang Hebei 050000, China
  • Received:2019-12-13 Revised:2020-02-24 Online:2020-07-10 Published:2020-05-13
  • Supported by:
    This work is partially supported by the Natural Science Foundation of Hebei (F2017202145),the Hebei Science and Technology Information Institute Open Project (HH1920).

基于改进残差结构的肺结节检测方法

石陆魁1, 马红祺1, 张朝宗2, 樊世燕1   

  1. 1. 河北工业大学 人工智能与数据科学学院, 天津 300401;
    2. 河北省科学技术情报研究院, 石家庄 050000
  • 通讯作者: 石陆魁
  • 作者简介:石陆魁(1974-),男,河北邯郸人,教授,博士,CCF会员,主要研究方向:机器学习、图像处理;马红祺(1994-),女,河北秦皇岛人,硕士研究生,主要研究方向:深度学习、医学图像处理;张朝宗(1980-),男,河北肃宁人,硕士,主要研究方向:大数据分析;樊世燕(1980-),女,河北衡水人,博士,主要研究方向:机器学习。
  • 基金资助:
    河北省自然科学基金资助项目(F2017202145);河北省科学技术情报院开放项目(HH1920)。

Abstract: In order to solve the problems of high computing cost and over-fitting of the model caused by complicated network structure in pulmonary nodule detection method, an improved residual network structure combining deep separable convolution and pre-activation was proposed. And the proposed network structure was applied to a pulmonary nodule detection model. Based on the target detection network Faster R-CNN, with U-Net coder-decoder structure adopted, the deep separable convolution and pre-activation operations were used by the model to improve the 3D residual network structure. Firstly, with the use of deep separable convolution, the complexity and computing cost of the model were reduced. Then, the regularization of the model was improved by introducing the pre-activation operation, and the phenomenon of overfitting was alleviated. Finally, the rectangular convolution kernel was used to expand the receptive field of the convolution operation on the premise that the computing cost of the model was slightly increased, so as to effectively take into account both the global and local characteristics of the pulmonary nodules. On the LUNA16 dataset, the proposed method has the sensitivity of 96.04%, and the Free-response area under the Receiver Operating Characteristic curve (FROC) score of 83.23%. The experimental results show that the method improves the sensitivity of pulmonary nodule detection, effectively reduces the average number of false positives in the detection results, and improves the detection efficiency. This proposed method can effectively assist radiologists in detecting pulmonary nodules.

Key words: pulmonary nodule detection, U-Net, Faster R-CNN, deep separable convolution, pre-activation

摘要: 针对肺结节检测方法中网络结构复杂所导致的模型计算量大、过拟合的问题,提出了一种结合深度可分离卷积和预激活的改进残差网络结构,将提出的网络结构应用于肺结节检测模型。该模型以目标检测网络Faster R-CNN为基础,采用U-Net编码解码器结构,利用深度可分离卷积和预激活操作改进了三维残差网络结构。首先,通过使用深度可分离卷积,模型复杂度和计算量大幅度降低;其次,通过使用预激活,模型的正则化得到改善,缓解了过拟合现象;最后,采用矩形卷积核在少量增加模型计算量的前提下扩大了卷积操作的感受野,有效地兼顾了肺结节的全局和局部特征。在LUNA16数据集上的检测中所提方法的灵敏度为96.04%,无限制接收者操作特征曲线下面积(FROC)得分为83.23%。实验结果表明:该方法提高了肺结节检测的灵敏度,又有效降低了检测结果的平均假阳性个数,同时提高了检测效率,可有效辅助放射科医师对肺结节进行检测。

关键词: 肺结节检测, U-Net, Faster R-CNN, 深度可分离卷积, 预激活

CLC Number: