摘要
针对现有视网膜血管分割技术存在视网膜血管分割精度不高和病灶区域误分割的问题,提出对U型网络改进,结合密集空洞卷积(dense atrous convolution,DAC)模块与残差挤压激励(residual squeeze and excitation,RSE)模块的视网膜血管分割模型(DACRSE-Unet)。该模型采用改进集成随机失活块(DropBlock)的残差结构,不仅可以构建深层网络来提取更复杂的血管特征,还可以有效缓解过拟合;此外,为了进一步提高网络的表达能力,在改进残差块的基础上引入挤压激励模块(squeeze and excitation,SE);同时,为获取血管更多的上下文信息,在模型中引入DAC模块来实现对视网膜血管的精准分割;最后,在不同数据集上进行验证。结果表明,DACRSE-Unet模型的接受者操作特性曲线下面积分别为0.9869和0.9964,灵敏度分别为0.8226和0.8779,准确率分别为0.9692和0.9830,整体分割效果比其他模型更好。
As existing retinal vascular segmentation technologies have such problems as low accuracy of retinal vascular segmentation and mis-segmentation of lesion areas,a modified retinal vessel segmentation algorithm for U-Net was proposed,and a DACRSE-Unet was designed that combines dense atrous convolution(DAC)module and residual squeeze and excitation(RSE)module.The network adopts the residual structure of improved integrated DropBlock,which can not only build a deep network to extract more complex vascular features,but also effectively alleviate overfitting.In addition,in order to further improve the expression ability of the network,a squeeze and excitation(SE)module is introduced on the basis of improving the residual block.At the same time,in order to obtain more contextual information about blood vessels,a DAC module is introduced in the middle of the network to achieve precise segmentation of retinal blood vessels.The proposed algorithm was validated on different datasets,with area under receiver operating characteristic curve of 0.9869 and 0.9964,sensitivity of 0.8226 and 0.8779,accuracy of 0.9692 and 0.9830,respectively.The overall segmentation effect is better than other algorithms.
作者
徐艳
张乾
吕义付
XU Yan;ZHANG Qian;LYU Yifu(School of Data Science and Information Engineering,Guizhou Minzu University,Guiyang 550025,China;Guizhou Key Laboratory of Pattern Recognition and Intelligent System,Guiyang 550025,China;Office of Academic Affairs,Guizhou Minzu University,Guiyang 550025,China)
出处
《湖北民族大学学报（自然科学版）》
CAS
2023年第3期360-367,共8页
Journal of Hubei Minzu University:Natural Science Edition
基金
贵州民族大学校级科研项目(GZMUZK[2021]YB23)。
关键词
U型网络
视网膜血管
图像分割
残差挤压激励模块
注意力机制
密集空洞卷积模块
U-Net
retinal blood vessels
image segmentation
residual squeeze and excitation module
attention mechanism
dense atrous convolution module
