基于改进HED网络的视网膜血管图像分割

Retinal Vascular Image Segmentation Based on Improved HED Network

视网膜血管的自动分割在糖尿病和高血压等疾病的诊断中起着重要作用。针对现有算法在细小血管和病变区域血管分割能力不足的问题,提出了一种基于改进整体嵌套边缘检测(HED)网络的视网膜血管分割算法。首先,采用了一种残差可变形卷积块代替普通卷积块,增强模型捕获血管形状和尺寸的能力;其次,采用扩张卷积层取代原有的池化层,用以保留血管特征的空间位置信息;最后,使用具有底部短连接结构的HED网络框架对预训练的网络进行特征提取和融合,使得模型可以更好地将骨干网络所提取的视网膜图像中血管的高级结构信息与低级细节信息相融合。通过在DRIVE(Digital Retinal Images for Vessel Extraction)和STARE(Structured Analysis of the Retina)数据集上进行验证,所提网络的灵敏度分别达到了81.75%和80.68%,特异性分别达到了97.67%和98.38%,准确性分别达到了95.44%和96.56%,受试者工作特征曲线(ROC)的曲线下面积(AUC)分别达到了98.33%和98.12%,实现了优于其他先进方法的综合分割性能。

Automated segmentation of retinal blood vessels plays an important role in the diagnosis of diseases such as diabetes and hypertension.Existing algorithms have insufficient ability to segment blood vessels into small blood vessels and lesions.In this paper,a retinal vessel segmentation method based on an improved holistically nested edge detection(HED)network is proposed to solve the segmentation problem.In the proposed method,firstly a residual deformable convolution block is used instead of the ordinary convolution block to enhance the ability of the model to capture the shape and size of the blood vessel;Subsequently,the original pooling layer is replaced by a dilated convolution layer to preserve the spatial locations of blood vessels;finally,an HED network framework with a short connection structure at the bottom is used for feature extraction and fusion of pre-trained networks,in which the model can better fuse the high-level structural information of the blood vessels and low-level details of the blood vessels in the retinal image extracted by the backbone network.By verifying the digital retinal images for vessel extraction(DRIVE)and the structured analysis of the retina(STARE)datasets,the sensitivities are 81.75%and 80.68%,the specificities are 97.67%and 98.38%,the accuracies are 95.44%and 96.56%,and the area under curve(AUC)of receiver operating(ROC)are 98.33%and 98.12%,respectively.The proposed method achieves comprehensive segmentation performance,which is superior to that of other advanced methods.