Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (2): 191-197.DOI: 10.3778/j.issn.1002-8331.1910-0340

Previous Articles     Next Articles

Research of Segmentation and Detection of Liver Metastases from CT Images Based on Deep Network

LI Jiasheng, GUO Shuxu, ZHANG Lei, ZHENG Shuang, ZHANG Huimao, QIU Yunhai, LI Xueyan   

  1. 1.College of Electronic Science and Engineering, Jilin University, Changchun 130012, China
    2.Department of Radiology, The First Hospital of Jilin University, Changchun 130021, China
  • Online:2021-01-15 Published:2021-01-14



  1. 1.吉林大学 电子科学与工程学院,长春 130012
    2.吉林大学第一医院 放射科,长春 130021


Evaluation of liver tumors is an important step in the clinical diagnosis and treatment of liver metastases from colorectal cancer. In order to complete the automatic segmentation and detection of liver tumors in abdominal CT images, an improved cascaded deep learning network is proposed. The cascaded network uses U-Net and Mask R-CNN models to perform segmentation and detection tasks, respectively. First, the U-Net model is trained as the first layer of the cascaded network to segment the liver organ as the Region of Interest(ROI). Then, the morphological activity contour extraction is performed for the ROI region. Finally, the U-Net model and Mask R-CNN are used as the second layer of the cascaded network, the task of accurately segmenting and detecting liver tumors in the ROI is completed, respectively. The experimental results show that the average Dice coefficient of liver metastasis segmentation for the cascade U-Nets model is 74%, the Dice coefficient of the tumor instance segmentation of Mask R-CNN is 67% (confidence is 95%), and the mean Average Precision(mAP) is 88%.

Key words: deep learning, medical image segmentation, object detection;cascaded network


肝脏肿瘤的评估是结直肠癌肝转移临床诊疗的重要步骤。为了完成腹部CT影像中的肝脏肿瘤自动分割和检测任务,提出一种改进的级联深度学习网络。级联网络采用U-Net和Mask R-CNN模型分别完成分割和检测任务。训练U-Net模型作为级联网络的第一层来分割肝脏器官作为感兴趣区域(ROI);针对ROI区域进行形态学活动轮廓提取;使用U-Net模型和Mask R-CNN模型作为级联网络的第二层分别完成精准分割和检测ROI内肝脏肿瘤的任务。实验结果表明,对于级联U-Net模型的肝脏转移瘤分割平均Dice系数为74%;Mask R-CNN的肿瘤实例分割Dice系数为67%(置信度为95%),均值平均精度(mAP)为88%。

关键词: 深度学习, 医学影像分割, 目标检测, 级联网络