基于改进R2U-Net型网络的视网膜血管图像分割

Image Segmentation of Retinal Blood Vessels Based on Improved R2U-Net Network

眼底视网膜血管的分割能够更有效地帮助医生诊断病情,但人工诊断费时耗力,传统的眼底图像血管分割技术的准确率和精度又不能达到理想状态,因此提出了基于R2U-Net的多尺度特征融合注意力网络——R2MAFF-Net。为了解决U-Net深度不够、上下层之间特征连接不密切及信息获取不完全等问题,将循环残差空洞卷积结构作为模型的编码部分,并将卷积层中的dropout替换成Dropblock,增强网络的抗拟合性能。在连接层中融入多尺度注意力机制,提高了血管和背景的对比度,改善了分割效果。在最底层使用多尺度残差特征金字塔模块,不同内容和不同尺度的特征融合使整个模型能够及时察觉各尺度变化。采用混合损失函数来提高模型的分割性能。该算法在公开数据库DRIVE和STARE的准确率达98.13%,98.20%,较原始的U-Net算法有一定的提升。

The segmentation of retinal blood vessels can help doctors diagnose the disease more effectively,but manual diagnosis takes more time and labor,and the accuracy and precision of traditional retinal vessel segmentation technology cannot reach the ideal state.Therefore,a multi-scale feature fusion attention network based on R2U-NET——R2MAFF-NET is proposed.In order to solve the problems of insufficient U-NET depth,incomplete connection of features between upper and lower layers,and incomplete information acquisition,the recurrent residual dilated convolution structure is taken as the coding part of the model,and dropout is replaced by Dropblock in the convolution layer to enhance the anti-fitting performance of the network.The multi-scale attention mechanism is integrated into the connection layer to improve the contrast between blood vessels and background and improve the segmentation effect.Multi-scale residual feature pyramid module is used at the bottom of the network,and the feature fusion of different contents and different scales enables the whole model to detect the changes of various scales in time.The mixed loss function is used to improve the segmentation performance of the model.The accuracy of the algorithm in public database DRIVE and STARE is 98.13%and 98.20%,which is better than the original U-Net algorithm.