低尺度血管检测在视网膜血管分割中的应用

视网膜图像分析成为目前诊断多种疾病非侵入的主要方式,其中血管的提取是最重要的一步。监督学习的方法在血管提取上有很好的效果,为了进一步提高检测的精度,提出了低尺度血管检测(LVD)算法。该网络除了有一个提取输入原尺度下特征的子网络外,还增加了两个低尺度下的特征提取子网络,并将低尺度下的单一输出融合原尺寸下的特征,降维后得到最后的输出结果。考虑到眼底血管结构特性,在LVD中设计了具有较深层数和较少参数的非对称固定深度子网络(ADS)。在公共的数据库DRIVE中进行测试,仅采用彩色眼底图像的绿色分量和B-COSFIRE滤波响应作为特征输入,其敏感性、特异性、准确率以及AUC指标分别为0.819 2、0.984 2、0.969 5、0.978 2,达到了先进水平。

Application of Low-Scales Vessel Detection in Retinal Vessel Segmentation

Retinal image analysis has become the main non-invasive way to diagnose many diseases, and the extraction of blood vessels is the most important step. Supervised learning method has a good effect on blood vessel extraction. In order to further improve the accuracy of detection, a low-scales vessel detection(LVD) algorithm is proposed. In addition to a sub-network for extracting features in the original scale, two sub-networks for extracting features in the low scale are added, and the single output in the low scale is fused with the features in the original size, and the final output result is obtained after dimensionality reduction. Considering the structural characteristics of fundus vessels, an asymmetric depth-fixed sub-network(ADS) with deep layers and fewer parameters is designed in LVD. Tested in the public database DRIVE, only the green component of color fundus image and B-COSFIRE filter response are used as feature input. Its sensitivity, specificity, accuracy and AUC index are 0.8192, 0.9842,0.9695 and 0.9782, respectively, which reach the advanced level.