文章摘要

文章信息/Info

Title:: Automatic extraction of carotid intima-media by ultrasound based on boundary saliency

作者:: 杨继锋1#; 韦浩1#; 熊飞2; 黄庆华3; 李乐4; 周光泉1△; （1.东南大学生物科学与医学工程学院，南京 210096；2.深圳市德力凯医疗设备股份有限公司，深圳 518132；3．西北工业大学光电与智能研究院，西安 710072；4．西北工业大学医学研究所，西安 710072）

Author(s):: YANG Jifeng1; WEI Hao 1; XIONG Fei 2; HUANG Qinghua 3; LI Le 4; ZHOU Guangquan 1; （1.School of Biological Science and Medical Engineering, Southeast University, Nanjing 210096, China; 2. Shenzhen Delica Medical Equipement Co.，Ltd, Shenzhen 518132, China; 3. School of Artificial Intelligence, OPtics and ElectroNics (iOPEN), Northwestern Polytechnical University, Xi′an 710072, China; 4. Institute of Medical Research, Northwestern Polytechnical University, Xi′an 710072）

Keywords:: Image segmentation; Carotid intima-media; Stripe attention; Auto-encoder; Cardio-cerebrovascular

摘要:: 为进一步提高超声颈动脉内中膜提取和测量的准确性,本研究基于U-Net模型提出了改进的分割网络，以实现对颈动脉内中膜的精准提取。首先，在网络中加入条形注意力模块，利用先验形状和解剖信息以解决传统卷积感受野受限的问题；此外，结合后处理细化模块以更好地减少图像中噪声和伪影干扰，通过从内中膜的固有膜形状特征中学习，从而实现校正估计误差。在采集的1 000张颈动脉血管超声图像数据库中进行测试，分割Dice达到0.932，内中膜厚度的平均误差为0.914个像素。本研究有望为动脉疾病的自动分析提供重要的参考依据。

Abstract:: In order to further improve the accuracy of intima-media extraction and measurement by ultrasound, we proposed an improved segmentation network based on U-Net model to achieve accurate extraction of carotid intima-media. Firstly, the stripe attention module was added to the network to solve the problem of traditional convolutional restricted receptive field by using prior shape and anatomical information. In addition, by combining the post processing refinement module. The interference of noise and artifacts in the image was reduced better, and the estimation error was corrected by learning from the intrinsic film shape features of the inner and middle film. The test was carried out in the database of 1 000 carotid artery ultrasound images collected. The segmentation Dice reached 0.932, and the average error of the thickness of the inner and media membranes was 0.914 pixels. This research is expected to provide important reference value for the automatic analysis of arterial diseases.