Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (18): 163-171.DOI: 10.3778/j.issn.1002-8331.2205-0134

• Graphics and Image Processing • Previous Articles     Next Articles

Segmentation Method of Retinal Vessels in OCTA with Attention Mechanism

CUI Shaoguo, WEN Hao, ZHANG Yunan, TANG Yibo, DU Xing   

  1. School of Computer and Information Sciences, Chongqing Normal University, Chongqing 401331, China
  • Online:2023-09-15 Published:2023-09-15

带有注意力机制的OCTA视网膜血管分割方法

崔少国,文浩,张宇楠,唐艺菠,杜兴   

  1. 重庆师范大学 计算机与信息科学学院,重庆 401331

Abstract: Retinal vascular segmentation is a key step in the process of intelligent aided diagnosis. Because the ends of the vessels are small and easy to be confused with the background, it is difficult to segment accurately. To solve this problem, an efficient retinal vessel segmentation algorithm based on deep separable convolution and block attention mechanism is proposed in this paper. Firstly, the convolution layer with step size of 2 is used to replace the maximum pooling layer for feature screening. Then, in order to reduce the complexity of the network, the general convolution operation is replaced by the depth separable convolution operation. Finally, the attention mechanism is introduced to learn important features to accurately classify the pixels of the OCTA image. The proposed method is tested on the 2020 public dataset OCTA-500. The results show that the F1, mIoU, Se, Sp, Acc and Pre are 80.01%, 81.18%, 84.39%, 96.41%, 95.32% and 76.24% respectively. Compared with the U-net, the parameters and FLOPs of this method are also significantly reduced, only 19.2% and 16.5% of the U-net.

Key words: deep learning, medical image processing, retinal vascular segmentation, depth separable convolution, block attention mechanism

摘要: 视网膜血管分割是智能辅助诊断过程中的关键步骤。由于血管末端细小且易与背景混淆,导致很难精确分割。针对此类问题,提出一种基于深度可分离卷积与块注意力机制的高效视网膜血管分割算法。使用步长为2的卷积层代替最大池化层进行特征筛选;为了降低网络的参数量,用深度可分离卷积代替常规的卷积;引入注意力机制学习重要特征,对OCTA视网膜血管图像的像素进行精确分类。将该方法在2020版的公开数据集OCTA-500上进行充分实验。结果表明,该方法在分割性能指标F1、mIoU、Se、Sp、Acc和Pre上分别达到了80.01%、81.18%、84.39%、96.41%、95.32%和76.24%;和基准方法U-net相比,该方法的参数量和FLOPs也显著降低,分别只有U-net的19.2%和16.5%。

关键词: 深度学习, 医学图像处理, 视网膜血管分割, 深度可分离卷积, 块注意力机制