数据与计算发展前沿 ›› 2024, Vol. 6 ›› Issue (3): 28-40.

CSTR: 32002.14.jfdc.CN10-1649/TP.2024.03.003

doi: 10.11871/jfdc.issn.2096-742X.2024.03.003

• 专刊：智慧医疗前沿与进展（下） • 上一篇    下一篇

基于外轮廓识别与内部像素分类的双阶段血管内超声影像分割与术中显示

于春宇1(),郭云涛2,王洪凯2,*()   

  1. 1.大连理工大学附属大连市中心医院，介入导管室，辽宁 大连 116021
    2.大连理工大学，医学部，辽宁 大连 116024
  • 收稿日期:2023-11-02 出版日期:2024-06-20 发布日期:2024-06-21
  • 通讯作者: *王洪凯（E-mail: wang.hongkai@dlut.edu.cn）
  • 作者简介:于春宇，大连理工大学附属大连市中心医院，主管技师，毕业于大连医科大学，从事介入导管室技师工作10余年。完成操作并解析腔内影像（IVUS、OCT）图像2000余例，参与编辑SCI文章3篇（最高影响因子IF=11）。2021年大连市医学重点专科“登峰计划”自主立项课题负责人1项。国家实用新型专利1项。
    本文中负责关键帧的选择，IVUS影像的阅读指导，实验结果分析。
    YU Chunyu, supervising technician, who graduated from Dalian Medical University, has been working as an interventional catheterization laboratory technician for more than 10 years. He has operated and analyzed more than 2,000 cases of intracavitary imaging (IVUS, OCT), and has participated in editing 3 SCI articles (highest IF=11). 2021 Dalian Medical Key Specialty "Peak Program" independent project leader. One national utility model patent.
    In this paper, he is responsible for the selection of key frames, the reading guidance of IVUS images, and the analysis of experimental results.
    E-mail: 36695194@qq.com|王洪凯，大连理工大学医学部，教授，博士生导师。主要研究方向为医学影像大数据统计建模分析，近年来在医学影像领域的顶级期刊以及Nature系列子刊发表论文多篇。担任了多个学术组织和机构的委员职务，包括全球华人学者医学图像计算青年研讨会主席、国际数字医学联盟委员、中国解剖学会断层影像解剖学分会副主任委员等。
    本文中负责提供研究指导，论文修订。
    WANG Hongkai is a professor and doctoral supervisor at the Department of Medicine, Dalian University of Technology. His main research interests include statistical modeling and analysis of big data in medical imaging, and he has published many papers in top journals in the field of medical imaging as well as the Nature series of subjournals in recent years. He has served as a member of several academic organizations and institutions, including the chair of the Global Youth Workshop on Medical Image Computing for Chinese Scholars, a member of the International Federation for Digital Medicine, and the vice-chairman of the Anatomical Branch of Tomographic Imaging of the Chinese Anatomical Society, etc. He is also a member of the International Federation for Digital Medicine.
    In this paper, he is responsible for providing research guidance and paper revision.
    E-mail: wang.hongkai@dlut.edu.cn
  • 基金资助:
    国家自然科学基金面上项目(81971693);大连理工大学-大连市中心医院医工交叉联合基金(DUT22YG229);大连理工大学-大连市中心医院医工交叉联合基金(DUT22YG205);大连市中心医院大连市医学重点专科“登峰计划”科研立项基金(2022ZZ229)

Two-Stage Segmentation and Intraoperative Visualization of Intravascular Ultrasound Images Based on Contour Recognition and Internal Pixel Classification

YU Chunyu1(),GUO Yuntao2,WANG Hongkai2,*()   

  1. 1. Interventional Catheterization Laboratory, Dalian Municipal Central Hospital affiliated with Dalian University of Technology, Dalian, Liaoning 116021, China
    2. School of Medicine, Dalian University of Technology, Dalian, Liaoning 116024, China
  • Received:2023-11-02 Online:2024-06-20 Published:2024-06-21

RichHTML

9

PDF

264

摘要：

【目的】血管内超声成像常被用于术中观察冠状动脉的狭窄状况与粥样硬化病灶危险程度。本文研发了针对血管内超声影像的病灶智能分割算法，重点解决噪声和伪影对分割精度的影响，并提出便于医生在术中观察的分割结果显示方式。【方法】本文提出了双阶段分割模型，首先自动识别血管内腔边轮廓以排除血管外噪声干扰，然后聚焦于血管内病灶组织的像素分类。本方法充分考虑了时序图像序列在时间上的关联性，通过多通道输入的U-Net提升相邻帧之间的分割一致性。为便于术中观察，根据网络输出的像素概率进行了概率化显示。【结果】在20例时序影像测试集上，本方法对纤维、钙化、脂质与超声衰减分割的平均Dice系数指标分别为0.90、0.93、0.80和0.95。对比实验证明外轮廓识别有助于排除外部噪声干扰，提升内部病灶分割的完整性。本方法以多通道方式输入时序图像可以有效提高时间维度上的分割一致性。临床医生验证肯定了本方法的概率化显示方式有助于术中直观了解病灶分布状况。【结论】本方法通过对血管内超声图像进行分割和可视化展示，更准确、完整、直观地评估并展现血管内病灶成份分布情况，为冠脉介入手术提供了智能分析和可视化展示的支持。

关键词: 血管内超声, 外轮廓识别, 像素标记, 动脉粥样硬化病灶分割

Abstract:

[Objective] Intravascular ultrasound imaging is often used for intraoperative observation of coronary artery stenosis and atherosclerotic plaque risk. In this paper, we develop an intelligent segmentation algorithm for intravascular ultrasound imaging, which focuses on solving the effects of noise and artifacts on the segmentation accuracy, and proposes a way to display the segmentation results that is easy for doctors to observe during the operation. [Methods] In this paper, a two-stage segmentation model is proposed, which first automatically identifies the luminal edge contour of the intravascular vessel to exclude the extravascular noise interference, and then focuses on the pixel classification of the intravascular plaque tissue. This method fully considers the temporal correlation of time-sequenced image sequences and enhances the segmentation consistency between neighboring frames by U-Net with multi-channel input. For intraoperative observation, a probabilistic display is performed based on the pixel probabilities output by the network. [Results] On a test dataset of 20-case time-series images, the average Dice coefficient indexes of fiber, calcification, lipid, and ultrasound attenuation segmentation by the present method are 0.90, 0.93, 0.80, and 0.95, respectively. Comparison experiments prove that external contour recognition helps to exclude external noise interference and enhance the completeness of the segmentation of internal lesions. This method can effectively improve the segmentation consistency in the time dimension by inputting time series images in a multi-channel manner. Clinicians have confirmed that the probabilistic display of this method is helpful for intraoperative visualization of lesion distribution. [Conclusions] By segmenting and visualizing intravascular ultrasound images, this method can more accurately, completely, and intuitively assess and display the distribution of intravascular lesion components, and provide intelligent analysis and visualization support for coronary interventional procedures.

Key words: intravascular ultrasound, outline recognition, pixel labeling, atherosclerotic plaque segmentation