当前的分割算法在处理眼底图像时存在的主要问题是特征细节的丢失。由于眼底图像中的血管形态学结构复杂多变,且受到伪影、噪声和病灶等多种因素的干扰,给分割过程带来了极大的困难。本文提出基于注意力机制以及双路上采样的眼底血管分...当前的分割算法在处理眼底图像时存在的主要问题是特征细节的丢失。由于眼底图像中的血管形态学结构复杂多变,且受到伪影、噪声和病灶等多种因素的干扰,给分割过程带来了极大的困难。本文提出基于注意力机制以及双路上采样的眼底血管分割算法(Self-Attention Dual Path Upsampling U-Net, SDU-Net),该算法引入两个模块,旨在提升视网膜血管分割的效率和准确率。注意力模块(Self-Attention, SEM),能够充分捕捉图像的上下文信息,减少信息损失,提高图像特征提取的准确性。同时,采用双路上采样(Dual Path Upsampling, DPUS)模块提高图像分辨率,对空间和信道信息进行补偿。在DRIVE/CHASE_DB1数据集上的实验表明,SDU-Net在灵敏度、特异性、AUC和F1-score指标方面优于其他方法。与原始U-Net相比,在DRIVE数据集上的实验中,Se提高了3.99%,Sp提高了0.3%,AUC提高了2.03%,F1提高了3.61%。在CHASE_DB1数据集上的实验中,Se提高了4.00%,Sp提高了1.06%,AUC提高了2.03%,F1提高了3.61%。这些结果表明,SDU-Net在视网膜血管分割任务中具有显著的优势。The main problem with current segmentation algorithms in processing fundus images is the loss of feature details. Due to the complex and varied vascular morphological structures in fundus images, as well as the interference of various factors such as artifacts, noise, and lesions, the segmentation process is greatly difficult. This article proposes a Self-Attention Dual Path Upsampling U-Net (SDU-Net) algorithm for retinal vessel segmentation based on attention mechanism and dual path upsampling. The algorithm introduces two modules to improve the efficiency and accuracy of retinal vessel segmentation. Self-Attention (SEM) module can fully capture the contextual information of images, reduce information loss, and improve the accuracy of image feature extraction. At the same time, the Dual Path Upsampling (DPUS) module is used to improve image resolution and compensate for spatial and channel information. Experiments on the DRIVE/CHASE_DB1 dataset have shown that SDU-Net outperforms other methods in sensitivity, specificity, AUC, and F1-score metrics. Compared with the original U-Net, the experiment on the DRIVE dataset showed an increase of 3.99% in Se, 0.3% in Sp, 2.03% in AUC, and 3.61% in F1. In the experiment on the CHASE_DB1 dataset, Se increased by 4.00%, Sp increased by 1.06%, AUC increased by 2.03%, and F1 increased by 3.61%. These results indicate that SDU-Net has significant advantages in retinal vessel segmentation tasks.展开更多
文摘当前的分割算法在处理眼底图像时存在的主要问题是特征细节的丢失。由于眼底图像中的血管形态学结构复杂多变,且受到伪影、噪声和病灶等多种因素的干扰,给分割过程带来了极大的困难。本文提出基于注意力机制以及双路上采样的眼底血管分割算法(Self-Attention Dual Path Upsampling U-Net, SDU-Net),该算法引入两个模块,旨在提升视网膜血管分割的效率和准确率。注意力模块(Self-Attention, SEM),能够充分捕捉图像的上下文信息,减少信息损失,提高图像特征提取的准确性。同时,采用双路上采样(Dual Path Upsampling, DPUS)模块提高图像分辨率,对空间和信道信息进行补偿。在DRIVE/CHASE_DB1数据集上的实验表明,SDU-Net在灵敏度、特异性、AUC和F1-score指标方面优于其他方法。与原始U-Net相比,在DRIVE数据集上的实验中,Se提高了3.99%,Sp提高了0.3%,AUC提高了2.03%,F1提高了3.61%。在CHASE_DB1数据集上的实验中,Se提高了4.00%,Sp提高了1.06%,AUC提高了2.03%,F1提高了3.61%。这些结果表明,SDU-Net在视网膜血管分割任务中具有显著的优势。The main problem with current segmentation algorithms in processing fundus images is the loss of feature details. Due to the complex and varied vascular morphological structures in fundus images, as well as the interference of various factors such as artifacts, noise, and lesions, the segmentation process is greatly difficult. This article proposes a Self-Attention Dual Path Upsampling U-Net (SDU-Net) algorithm for retinal vessel segmentation based on attention mechanism and dual path upsampling. The algorithm introduces two modules to improve the efficiency and accuracy of retinal vessel segmentation. Self-Attention (SEM) module can fully capture the contextual information of images, reduce information loss, and improve the accuracy of image feature extraction. At the same time, the Dual Path Upsampling (DPUS) module is used to improve image resolution and compensate for spatial and channel information. Experiments on the DRIVE/CHASE_DB1 dataset have shown that SDU-Net outperforms other methods in sensitivity, specificity, AUC, and F1-score metrics. Compared with the original U-Net, the experiment on the DRIVE dataset showed an increase of 3.99% in Se, 0.3% in Sp, 2.03% in AUC, and 3.61% in F1. In the experiment on the CHASE_DB1 dataset, Se increased by 4.00%, Sp increased by 1.06%, AUC increased by 2.03%, and F1 increased by 3.61%. These results indicate that SDU-Net has significant advantages in retinal vessel segmentation tasks.