基于SENet注意力机制和深度残差网络的腹部动脉分割

Abdominal Artery Segmentation Based on SENet Attention Mechanism and Deep Residual Network

在医学诊断中,血管疾病的研究与治疗仍是影响人类健康的主要因素。由于人体腹部血管复杂且构造因人而异,这就对图像分割的研究以及临床应用带来了极大困难。所以,通过图像处理和深度学习等方法准确清晰地获取病人腹部动脉及其分支血管,在临床和术前诊断中发挥了重要作用。本文主要对腹部血管的大小、灰度、构造等基础医学知识进行学习,并深入研究了现有关于血管分割算法的优缺点。为解决深度卷积神经网络性能退化的问题,增强对目标信息的关注度并对不必要的特征信息进行抑制,本文提出一种基于squeeze-and-excitation networks(SENet)的注意力机制和深度残差网络的血管分割算法。并使用12例腹部CT数据进行实验验证:血管分割准确率可达90.48%,灵敏度、Dice、VOE、精确率分别为0.899 5、0.878 3、-0.199 8、0.910 4。因此,相比于传统方法,本实验所提方法具有更好的分割性能。

The study and treatment of vascular diseases in medical diagnosis remains a major factor affecting human health. Since human abdominal blood vessels are complex and their structure varies from person to person, this poses great difficulties for the research of image segmentation as well as clinical applications. Therefore, accurate and clear acquisition of a patient’s abdominal arteries and their branch vessels by methods such as image processing and deep learning plays an important role in clinical and preoperative diagnosis. Basic medical knowledge about the size, grayscale and configuration of abdominal blood vessels was learned, and the advantages and disadvantages of existing algorithms about blood vessel segmentation were studied in depth. In order to solve the performance degradation of deep convolutional neural networks, enhance the focus on target information and suppress unnecessary feature information, a vessel segmentation algorithm based on squeeze-and-excitation networks(SENet) and deep residual networks was proposed. The evaluation results using 12 cases of abdominal CT data show that the accuracy of vessel segmentation could reach 90.48%, and the sensitivity, Dice, VOE, and accuracy are 0.899 5, 0.878 3,-0.199 8 and 0.910 4, respectively. therefore, the proposed method in this experiment has better segmentation performance compared with the conventional method.