多尺度卷积核U-Net模型的视网膜血管分割方法

针对病变视网膜血管结构的计算机辅助诊断问题，提出了一种多尺度卷积核U-Net模型的视网膜血管分割方法.在U-Net模型基础上设计了融合Inception模块和最大索引值上采样方法的多尺度卷积神经网络结构.在网络训练阶段，采取旋转、镜像等操作进行数据集扩充，运用CLAHE算法进行图像预处理；训练后得到的双通道特征图，进行Softmax归一化；最后通过改进的代价损失函数对归一化结果迭代优化，得到完整的视网膜血管分割模型.实验结果表明，所提方法在DRIVE数据集上分割的准确率达到0.9694，灵敏性达到0.7762，特异性达到0.9835，比U-Net模型具有更优的分割效果和泛化能力，与其他现存方法相比具有一定的竞争力.

Retinal Blood Vessel Segmentation Method Based on Multi-scale Convolution Kernel U-Net Model

Aiming at the computer-aided diagnosis of diseased retinal vascular structure， a retinal blood vessel segmentation method of multi-scale convolution kernel U-Net model was proposed. Based on the U-Net model， a multi-scale convolutional neural network structure combining with the Inception module and the maximum index value upsampling method was designed. In the network training stage， operations such as rotation and mirroring were used to expand the data sets， and the CLAHE algorithm was used for image preprocessing. The dual-channel feature map obtained after training was normalized by Softmax. Finally， the normalized result was iteratively optimized by the improved cost loss function， then a complete retinal vessel segmentation model was obtained. Experimental results showed that the proposed method on the DRIVE data set achieved an accuracy of 0.9694， a sensitivity of 0.7762， and a specificity of 0.9835. The proposed method has better segmentation effect and generalization ability than the U-Net model， and shows its competitive results compared with other existing methods.