基于U-Net的多尺度视网膜血管分割方法

Multi-scale Segmentation of Retinal Vessels Based on U-Net

视网膜血管的形态和结构一直是高血压、冠心病、糖尿病等疾病的重要诊断指标之一,其检测和分割具有十分重要的意义。为了解决视网膜血管分割中,血管末梢缺失和细小血管断裂的问题,提出了一种基于U-Net改进模型的多尺度分割方法,通过在编码阶段和解码阶段之间采用增加卷积块的方式来保持对不同尺度下的特征提取,同时对增加的卷积块采用密集连接的方式解决由于网络加深带来的浅层特征缺失和梯度消失问题,从而增强模型的特征提取能力并提高分割性能。此外,采用Dice损失函数解决数据集中正负样本不均衡的问题。实验采用CHASE_DB1和DRIVE两个数据集进行训练和测试,通过与U-net、Residual U-net、Ladder-Net以及R2U-Net的对比表明,由于保留了多尺度的细节信息,该方法取得了更好的分割效果。实验证明,该方法能够有效提取健康视网膜图像和病变视网膜图像中的血管网络,能够较好地分割细小血管。

The morphology and structure of retinal vessels have always been one of the important diagnostic indicators of hypertension,coronary heart disease,diabetes and other diseases,and their detection and segmentation are of great significance.In order to solve the problems of loss of vascular endings and rupture of small vessels in the segmentation of retinal vessels,a multi-scale segmentation method based on the improved U-Net model is proposed,which maintains feature extraction at different scales by adding convolutional blocks between the encoding stage and the decoding stage.At the same time,the added convolution blocks are densely connected to solve the problems of shallow layer feature loss and gradient disappearance caused by the deepening of the network,so as to enhance the feature extraction ability of the model and improve the segmentation performance.In addition,Dice loss function is used to solve the problem of unbalanced positive and negative samples in the data set.Two data sets,CHASE_DB1 and Drive,were used for training and testing in the experiment.The comparison with U-Net,Residual U-Net,Ladder-Net and R2 U-Net shows that the proposed method has achieved better segmentation effect due to the retention of multi-scale details.Experimental results show that the proposed method can effectively extract the vascular network from healthy and pathological retinal images,and can better segment small vessels.