MS-UNet++:基于改进UNet++的视网膜血管分割

本文针对视网膜图像中细微血管特征提取困难导致其分割难度高等问题,提出了一种 基于端到端的神经网络嵌套视网膜血管分割模型算法(简称MS-UNet++),该算法选取了深度监督网络UNet++作为分割网络模型,提升特征的使用效率；引入MulitRes模块,改善低对比度环境下细小血管的特征学习效果,并在特征提取后加上SENet模块进行挤压和激励操作,从而增强特征提取阶段的感受野,提高目标相关特征通道的权重。基于DRIVE图像数据集的实验结果表明,该算法分割结果与真实结果之间的重叠率DICE值为83.64%,并交比IOU为94.83%,准确度ACC为96.79%,灵敏度SE为81.78%,较现有模型有一定的提升,可用于视网膜图像血管分割,为临床诊断提供辅助信息。

MS-UNet++:Retinal vessel segmentation method based on improved UNet++

In this paper,an end-to-end neural network-based nested retinal vesse l segmentation model algorithm (MS-UNet++) is proposed to solve the problem of h igh segmentation difficulty caused by the difficulty of extracting subtle vessel features in retinal images.The algorithm selects the deep supervision network UNet++ as the segmentation network model to improve the use efficiency of featur e.In our algorithm design,We choose to introduce MulitRes module for feature e xtraction to improve the feature learning effect of the small blood vessels in l ow contrast environment.And after feature extraction,we select to insert SENet module is added for squeezing and exciting operations,so as to enhance the rec eptive field in feature extraction stage and increasing the weight of target-re l ated feature channels.The experimental results based on DRIVE image data set sh ow that the overlap ratio DICE value between with the segmentation result and th e ground truth is 83.64%,the parallel ratio IOU is 94.83%,and the accuracy is 96.79%,and the sensitivity is 81.78%,compared with the existing algorithm mode ls,the results of our algorithm models has a certain improvement.As a result, our algorithm models can be used for retinal image vessel segmentation and provi de auxiliary information for clinical diagnosis.