Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (3): 207-214.DOI: 10.3778/j.issn.1002-8331.2008-0214

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Pneumothorax Segmentation Method Based on Improved U-net Network

YU Sheng, WANG Kangjian, HE Lingmin, XU Zhijie, WANG Xiuhui   

  1. 1.College of Information Engineering, China Jiliang University,Hangzhou 310018, China
    2.Key Laboratory of Electromagnetic Wave Information Technology and Metrology of Zhejiang Province, China Jiliang University, Hangzhou 310018, China
  • Online:2022-02-01 Published:2022-01-28

基于改进U-net网络的气胸分割方法

余昇,王康健,何灵敏,胥智杰,王修晖   

  1. 1.中国计量大学 信息工程学院,杭州 310018
    2.中国计量大学 浙江省电磁波信息技术与计量检测重点实验室,杭州 310018

Abstract: Pneumothorax is one of the common lung diseases. The existing X-ray pneumothorax detection methods mainly have two problems:first, pneumothorax is usually overlapped with the ribs, clavicle and other tissues, which is clinically missed. Second, the existing mainstream segmentation algorithms adopt single or double threshold strategies, resulting in inaccurate results. To solve the above problems, a novel pneumothorax segmentation method is proposed. The contrast limited adaptive histogram equalization is applied to chest radiographs to remove noise points and restore image details. Using the convolutional neural network layer with MBConvBlock as the encoder module to extract the abstract deep features in the image. Then the feature map is interpolated and reconstructed by the decoder to obtain the binary classification result of each pixel. The improved triple threshold strategy is adopted to better meet the results of actual medical scenarios. Compared with DeepLabV3+ and U-net, the Dice similarity coefficient value, accuracy rate and recall rate obtained by this method on the Siim-ACR Pneumothorax data set are 87.21%, 94.81% and 88.96%, respectively. The experimental results show that this method can make the segmentation of X-ray pneumothorax with high precision and fill the shortage of the present pneumothorax image segmentation.

Key words: pneumothorax segmentation, triple threshold strategy, MBConvBlock, SIIM-ACR Pneumothorax

摘要: 气胸是肺部常见疾病之一,目前已有的X线气胸检测方法主要存在两个问题:一是气胸通常与肋骨、锁骨等组织重叠,在临床上存在较大的漏诊情况;二是现有的主流分割算法采用单一或双重阈值策略,导致结果不准确。针对上述问题,提出了一种新颖的气胸分割方法。该方法对胸片进行对比度限制自适应直方图均衡化,去除噪点并还原图像细节;通过以MBConvBlock为编码器模块的卷积神经网络层提取图像中抽象的深层特征;通过解码器对提取到的特征映射进行插值重构得到每个像素的二分类结果;采取改进的三重阈值策略输出更满足实际医用场景的结果。该方法在SIIM-ACR Pneumothorax数据集上得到的Dice相似系数值、精确率和召回率分别为87.21%、94.81%和88.96%,相比DeepLabV3+和U-net等网络,在气胸分割取得了更好的性能。实验结果表明该方法能够使X线气胸分割具有较高的精度,填补了目前气胸X线图像分割领域的不足。

关键词: 气胸分割, 三重阈值策略, MBConvBlock, SIIM-ACR Pneumothorax