Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (22): 219-228.DOI: 10.3778/j.issn.1002-8331.2104-0189

• Graphics and Image Processing • Previous Articles     Next Articles

Enhanced Network for Ultrasound Breast Tumor Segmentation Based on U-Net

CHEN Xi, LIU Qi, DENG Xiaobo, HE Kechen, QUAN Meilin   

  1. 1.College of Electrical Engineering, Sichuan University, Chengdu 610065, China
    2.College of Biomedical Engineering, Sichuan University, Chengdu 610065, China
  • Online:2022-11-15 Published:2022-11-15

改进U-Net的超声乳腺肿瘤分割网络

陈曦,刘奇,邓小波,何柯辰,全美霖   

  1. 1.四川大学 电气工程学院,成都 610065
    2.四川大学 生物医学工程学院,成都 610065

Abstract: The conventional U-Net performs poorly in the segmentation of breast tumors in ultrasound images, due to the variability of breast tumors’ shape, the image shadows and the obscure boundary. To overcome these difficulties, this paper proposes an enhanced network, namely MultiMixU-Net. The network introduces the MultiMix block and the Respath into the conventional U-Net. The MultiMix block uses the atrous convolution path to strengthen the ability to discern the target from the background. To extract multi-scale feature information, the MultiMix block concatenates outputs from convolution layers in the atrous convolution path and fuses outputs from the normal convolution paths. The network also modifies the layout of the Respath. This modification can make the transformation of feature information from the contract path to the expanse path more efficient. The experiments on the public ultrasound breast tumor segmentation dataset show that the MultiMixU-Net outperforms other networks and has fewer parameters. Compared with the conventional U-Net, the MultiMixU-Net in all performance metrics has better results. IoU and DSC are improved by 0.1541 and 0.1273, respectively.

Key words: breast tumor segmentation, ultrasound images, U-Net, deep learning

摘要: 乳腺超声图像具有肿瘤大小形态多变、阴影较多、边界模糊等特点,经典U-Net的乳腺肿瘤分割结果与标注图像出入较大。对此,提出改进网络MultiMixU-Net。该网络在U-Net结构中引入MultiMix block以及Respath。MultiMix block通过空洞卷积通路提高网络区分目标以及背景的能力,并通过级联该通路中各卷积层输出,融合普通卷积通路的输出来提取多尺度特征信息。Respath的改进部署使网络中收缩路径与扩张路径之间对应特征信息的传递更加有效。该改进网络在公开的超声乳腺肿瘤分割数据集上进行了测试,实验表明,MultiMixU-Net分割结果优于其他网络且参数量较少。相较于U-Net,所提网络分割结果在所有评价指标上均有提升,其中IoU、DSC分别提升0.154 1、0.127 3。

关键词: 乳腺肿瘤分割, 超声图像, U-Net, 深度学习