一种基于超像素和生成对抗网络的视网膜血管分割方法

针对传统视网膜图像血管分割中部分血管轮廓粗糙、血管末梢和分支细节丢失等问题，提出 一种结合线性谱聚类超像素与生成对抗网络(Generative Adversarial Networks，GAN)的视网膜血管分割 方法。该方法首先对 GAN 进行改进，采用空洞空间金字塔池化模块的多尺度特征提取来提高 GAN 分 割精度，在获得视网膜血管分割图像后，利用线性谱聚类超像素分割的边缘贴合性高、轮廓清晰的特 点，将 GAN 输出图像映射到超像素分割图再对像素块进行分类，以达到分割的效果。仿真实验结果表 明，与传统的视网膜血管分割方法相比，该方法在灵敏度和准确性上有一定提升，轮廓边缘细节方面 有着更好的效果。

A Retinal Vessel Segmentation Method Based on Super-pixel and Generative Adversarial Networks

In order to solve rough contour of some blood vessels and the loss of vessel-perpherals and branches in traditional retinal vessel segmentation, a novel method forretinal vessel segmentation which combines linear spectral clustering super-pixel with generative adversarial networks (GAN) is proposed.The accuracy of segmentation is improved using the multi-scale features from atrousspatial pyramid pooling (ASPP) module with a modified GAN method. After the segmentation image is obtained, by utilizing the characteristics of high edge suitability and clear contour of linear spectral clustering super-pixel segmentation, the GAN output image was mapped to the super-pixel segmentation image. The segmentation was achieved by classifying the pixel clusters. The experimental results show that compared with the traditional retinal vessel segmentation method, the sensitivity and accuracy of the proposed method are improved, especially in the details of the contour edge.