电子学报

• •    

基于结构重参数化与多尺度深度监督的COVID-19胸部CT图像自动分割

刘金平, 吴娟娟, 张荣, 徐鹏飞   

  1. 湖南师范大学信息科学与工程学院,湖南长沙410006
  • 收稿日期:2022-04-07 修回日期:2022-11-22 出版日期:2023-01-28
    • 通讯作者:
    • 徐鹏飞
    • 作者简介:
    • 刘金平 男,1983年9月出生于湖南省邵阳市.现为湖南师范大学信息科学与工程学院教授、博士生导师,主要从事智能信息处理、过程检测、故障诊断等相关研究.E-mail: ljp202518@163.com
      吴娟娟 女,1996年5月出生于湖南省株洲市.目前在湖南师范大学信息科学与工程学院攻读硕士学位.研究方向为生物医学图像处理. E-mail: Ajuan0527@hunnu.edu.cn
      张荣 女,1999年12月出生于湖南省长沙市.目前在湖南师范大学信息科学与工程学院攻读硕士学位.研究方向为生物医学图像处理. E-mail: cpstzr@163.com
      徐鹏飞(通讯作者) 男,1977年9月出生于湖南省岳阳市.现为湖南师范大学信息科学与工程学院副教授,主要从事智能信息处理方面的研究. Email: xupf@hunnu.edu.cn
    • 基金资助:
    • 国家自然科学基金 (61971188)

Toward Automated Segmentation of COVID-19 Chest CT Images Based on Structural Reparameterization and Multi-Scale Deep Supervision

LIU Jin-ping, WU Juan-juan, ZHANG Rong, XU Peng-fei   

  1. College of Information Science and Engineering,Hunan Normal University,Changsha,Hunan 410081,China
  • Received:2022-04-07 Revised:2022-11-22 Online:2023-01-28
    • Corresponding author:
    • XU Peng-fei
    • Supported by:
    • National Natural Science Foundation of China (61971188)

摘要:

基于深度学习模型的胸部CT(Computed Tomography)图像自动分割有助于辅助医生诊疗.但随着网络宽度与深度的加深,网络训练困难且推理减慢.为提高隐藏层的学习能力,深度监督机制被用于网络训练.但以往的深度监督方法没有考虑模型中多尺度特征图的分层表示以及上采样对参与损失计算的特征图质量的影响.为加强隐藏层学习过程的直接性同时加快网络推理,本文提出一种结构重参数化与多尺度深度监督分割网络(Structural Reparameterization and Multi-Scale Deep Supervision Network,SR&MDS-Net),以实现COVID-19(COrona VIrus Disease 2019)胸部CT图像的高效准确分割.首先构建一种结构重参数化特征变异(Structure Reparameterized FeatureVariation,SRFV)模块将网络的训练与推理进行解耦,在提高模型表达能力的同时加快推理速度;然后,提出一种新颖的多尺度深度监督机制,以加强网络监督效果,提高网络性能.在公开的COVID-19胸部CT图像数据集上进行实验,SR&MDS-Net的灵敏度、特异性、准确率、Dice分别达到了91.5%、99.5%、72.8%、80.1%,与同类其他方法比较,具有更优的性能.

关键词: COVID-19, 医学图像分割, 深度学习, U-Net, 结构重参数化, 深度监督学习

Abstract:

Automatic segmentation of chest CT(Computed Tomography) images based on deep learning models is helpful to assist doctors in diagnosis and treatment. However, with the deepening of the network width and depth, the network training is difficult and the inference slows down. In order to improve the learning ability of the hidden layer, the deep supervision mechanism is used in network training. However, previous deep supervision methods did not considerthe layered representation of multi-scale feature maps in the segmentation model and the influence of up-sampling on the quality of feature maps involved in loss calculation. In order to improve the directness of the hidden layer learning process and speed up the reasoning. This article proposes a structural reparameterization and multi-scale deep supervision network(SR&MDS-Net) for the purpose of achieving the efficient and accurate segmentation results of lesions in COVID-19(COrona VIrus Disease 2019) pneumonia chest CT images. Structure reparameterized featurevariation(SRFV) module was constructed to decoupage the network training and reasoning, which improved the model expression ability and accelerated the reasoning speed. In order to strengthen the effect of network supervision and improve the performance of network, a novel multi-scale deep supervision mechanism is proposed. The experiment was carried out on the public COVID-19 pneumonia chest CT image dataset. The sensitivity, specificity, accuracy, and Dice of SR&MDS-Net reached 91.5%, 99.5%, 72.8%, and 80.1%, respectively, which had better performance compared with other similar methods.

Key words: COVID-19, medicalimage segmentation, deep learning, U-Net, structure reparameterization, supervised learning

中图分类号: